<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sebastian Thielke on Medium]]></title>
        <description><![CDATA[Stories by Sebastian Thielke on Medium]]></description>
        <link>https://medium.com/@SThielke?source=rss-63c996d9e768------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*wmMtKDfIepBYJko1o8lzfQ@2x.jpeg</url>
            <title>Stories by Sebastian Thielke on Medium</title>
            <link>https://medium.com/@SThielke?source=rss-63c996d9e768------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 02:47:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@SThielke/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[More Roads for Meaning]]></title>
            <link>https://medium.com/@SThielke/more-roads-for-meaning-1775b721fa03?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/1775b721fa03</guid>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[ontology]]></category>
            <category><![CDATA[context]]></category>
            <category><![CDATA[meaning]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Tue, 07 Jul 2026 17:53:09 GMT</pubDate>
            <atom:updated>2026-07-07T17:53:09.483Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BAB_PpcsU5UueGnq.png" /></figure><p>There is a thing every traffic engineer learns and every city forgets. You have a congested highway. You widen it. For a year or two it flows. Then it fills again, and the new equilibrium carries more cars, more slowly, than the old one did. The wider road did not absorb the demand. It manufactured it. Lower the cost of driving and you draw out every trip people were quietly choosing not to take, and you reshape where they live and work around the road you just built, so that the congestion comes back wearing a larger size.</p><p>This has a name. Induced demand. The lesson inside it is not about cars. It is that when you treat a structural problem as a capacity shortage, every unit of capacity you add feeds the thing you were trying to reduce. Congestion was never a quantity of road. It was a structure of incentives. More road cannot reach a problem that more road is the wrong shape to solve, and worse, the road changes the world around it so the old, walkable, transit served equilibrium becomes harder to return to with every lane you pour.</p><p>I want to put that idea next to what the field building agentic AI is doing right now, because they are running the same play, and they cannot see it, for the same reason the city cannot see the highway.</p><h3>The capacity reflex</h3><p>An AI system produces fluent output that turns out, under pressure, not to be about anything. Ask it a real follow up question and the fluency cracks, because there was never a link between its symbols and the world those symbols name. This is the actual problem, and it is old. A model of language built on the statistics of words captures the regularities that meaning leaves behind in text. It captures the shadow that meaning casts. It does not capture the thing casting the shadow, because that thing, the bond between a sign and what the sign is about, was never in the data to begin with. That bond is a relation between a symbol and the world outside it, and a formal system is closed over its own symbols, defined entirely by how they relate to each other. A relation reaching outside that closure is not the kind of property such a system can hold, no matter how much structure you add on the inside. The aboutness lives in people standing in situations, using words to mean things, and paying for getting them wrong.</p><p>So the system does not understand. What does the field do?</p><p>It adds capacity. If the symbol does not carry meaning on its own, surround it with structure until it seems to. Add a knowledge graph, so the relations between things are drawn out explicitly, nodes and labeled edges, here is what connects to what. Add retrieval, so the model has more surrounding material to lean on. Widen the context window, so more of the world can be held at once. Each of these is a lane. Each lowers the cost of producing output that looks like understanding. And each fills back up, because the gap between a notation and the world it points at is not a capacity shortage, and you cannot pave across it.</p><p>A knowledge graph does not hold meaning. It holds another notation for meaning. The edge labeled “owns” or “depends on” or “is a kind of” means something to the person reading the graph, who supplies the aboutness out of their own head, exactly as they did before the graph existed. Drawing the relation more neatly does not put the relation inside the structure. It moves the same ungrounded symbol from a sentence into a node and gives it a cleaner shape. The grounding still sits outside the system, in a human. The graph widened the road. The traffic, the need for a person to actually read the situation and say what it means, came right back, and now it comes back through a more elaborate and more expensive piece of infrastructure that everyone has agreed to treat as if it were the destination.</p><h3>The road that reshapes the city</h3><p>The deeper damage in induced demand is not that the highway fills. It is that the highway, once built, changes everything around it so the alternatives wither. Sprawl follows the interchange. Transit loses its riders. The corner shop dies and the big box at the off ramp lives. After twenty years the city is shaped so that driving is the only thing that works, and the road you built to relieve congestion has made a world that cannot function without it.</p><p>The knowledge graph does this too, and this is the part that should worry people who are betting on it. As an organization leans harder on richer notation to stand in for understanding, it rebuilds itself around the premise that meaning is stored. Roles get designed around maintaining the graph. Decisions get routed through what the graph can represent. The person who used to hold the meaning, who read the situation and recognized that this case is not like the others, becomes an inconvenience to the schema, then a bottleneck to be removed, then a line item that was cut two reorganizations ago. The graph does not merely fail to hold meaning. It quietly dismantles the one thing that was holding it, the same way the highway dismantles the neighborhood that made it unnecessary.</p><p>So when someone says a knowledge graph will give the agents context, hear it the way you would hear a transport minister promising that one more lane will finally fix the commute. The promise is sincere. The person making it is inside a worldview where capacity is the only lever there is. And the promise is wrong in the specific way that creates the appetite it claims to satisfy.</p><h3>Ontologies, or the master plan</h3><p>If the knowledge graph is another lane, the ontology is the master plan for all future traffic, and it fails in the way master plans fail, which is more completely than any single road.</p><p>A graph draws the relations that happen to hold. An ontology goes further. It declares what the categories are, formally, in advance. It is a specification of meaning itself, written down once, as the authoritative schema that everything afterward must conform to. This is the capacity reflex at its most ambitious, and its ambition is exactly its weakness. The more thoroughly you fix the meaning of your terms ahead of time, the more total the mismatch becomes the moment the world moves and your specification does not. An ontology is a definition of every situation, made by people not in any of them, frozen at the moment of writing, applied forever after to situations it was never a reading of.</p><p>What it cannot do is the one thing that matters. An ontology can define “customer” as a class with properties and relations. It cannot perform the act of recognizing that this particular party, here, today, in circumstances nobody anticipated, is or is not a customer in the sense that should govern what happens next. That recognition is a reading done by someone standing in the situation. The class is a noun sitting in the schema. The recognition is a verb, and it happens in a person, and the ontology has no place to put it, so it pretends the question was already settled when the schema was written. Most of the time the pretence holds, because most cases are ordinary. It fails exactly on the cases that were not anticipated, which are the only cases where you needed judgment in the first place.</p><p>And there is the governance turn, which makes the ontology worse than the graph. A graph that has gone stale is merely unhelpful. An ontology that has gone stale is binding. People are required to conform to it. So when the world moves and the schema does not, the live reading that would catch the drift now contradicts the official categories, and contradicting the official categories is a violation. The structure does not just miss the meaning. It forbids the act that would recover it. You have seen this in human form. The long tenured employee who recites policy, fluent and certain, whose answers cannot survive one honest follow up question, because they froze their reading of the job years ago and the world kept moving without them. An organization running on an ontology is that employee, rendered in software and made mandatory.</p><h3>We have run this experiment</h3><p>If that failure sounds hypothetical, it is not. None of this is new, and we already know how it goes.</p><p>Formal knowledge representation has a long history. The semantic web was going to make the world’s information machine understandable by giving everything an ontology and a web of typed relations. Expert systems were going to capture professional judgment as rules. Cyc spent decades and millions of hand authored assertions trying to encode common sense as a formal knowledge base. These were serious efforts by serious people, and they did not stall for lack of detail or lack of funding or lack of cleverness. They stalled on the same wall every time. No finite formal specification of meaning ever closes, because the act that closes it is a reading, and a reading is not the kind of thing that goes into a schema. You can always write one more axiom. The world will always present the case the axioms did not cover, and covering it requires someone to look and decide, which is the capability the whole project was trying to make unnecessary.</p><p>The current wave has bolted embeddings and language models onto these old structures and expects the combination to come out somewhere new. It will not, on the part that matters, because the addition does not touch the wall. A graph built by a language model is still a graph. A category proposed by a model and written into an ontology is still a frozen category. The model is fluent enough to hide the seam, which means the failure arrives later and more expensively than it used to, but it arrives in the same place. We are repeating a settled experiment with better production values and reading the better production values as a reason to expect a different result.</p><p>There is one answer to all of this that sounds like it escapes the trap, and it does not. The graph and the ontology are static, the objection runs, so of course they go stale, but a world model is alive. It updates continuously against reality, it carries causal structure, it has a present. Surely the dynamic version crosses the gap the frozen ones cannot. It does not, and for a reason that has nothing to do with how fast it updates. A model that updates against reality is still relating its own internal states to one another, tuning them so the next prediction lands closer to the last observation. What flows in is more observation. What changes is the arrangement of the model’s own symbols. The relation to the world outside it, the aboutness, is no more present after the update than before, because updating is an operation the system performs on its own interior, and the interior is exactly the place the link does not live. A live model is a more faithful map. It is still a map, and the thing it most needs to navigate is not on the territory it maps. I have made the fuller case elsewhere, in the piece on why world models will not give a system meaning, and it is the same wall this whole essay is about, met by a system that moves instead of one that sits still.</p><h3>Where the road actually goes</h3><p>None of this means a graph is useless or an ontology is foolish. A map is a fine thing. A schema is a fine thing. The error is never that these structures exist. The error is believing the structure holds the meaning, rather than holding a notation of it that a person still has to read.</p><p>There is a place for the frozen thing. It is downstream of the human who holds the meaning, not in place of them. The graph, the ontology, the document, the spec, these are all containers into which a person freezes their current reading so that a fast machine can run against it for a while. That is real and useful work. What makes it useful is precisely that it stays subordinate to the reading. The human looks at the world, decides what it means now, freezes that into the structure, and the agent runs on the frozen version until the world moves enough that the human freezes it again. The structure is the noun. The person is the verb. The system works when the noun is kept close to the verb by someone whose job is to keep reading.</p><p>The road analogy holds right up to its edge, and then it breaks in a way that matters. Congestion has a real answer inside the transport system. You can price the road, build the transit, change the zoning, and actually move the equilibrium, because congestion is a hard problem but not a categorical one. Meaning is different and harder. There is no rearrangement of notation, however clever, that turns a symbol into something that is about the world, because being about the world is not a property a notation can have. It is a property of a participant who lives in the world and bears the cost of reading it wrong. So with roads, the fix is elsewhere in the same system. With meaning, the fix is a different kind of thing entirely, and no amount of building inside the formal system substitutes for it.</p><p>That is the whole of it. The field is widening the road and counting the new lanes as progress, while the traffic it is trying to clear is the steady, irreducible need for someone to stand in the situation and say what it means. You cannot pave that need away. Every lane you add to avoid it draws more of it out, and quietly makes the one thing that could carry it, a person holding meaning, harder to keep on the payroll. The knowledge graph is induced demand for understanding. The ontology is the master plan that mandates yesterday’s understanding forever. And the experiment that shows where both of them end was run a generation ago, finished, and written down, and the field is funding the sequel without having read the first result.</p><p>This is part of Human in Meaning, the argument that meaning is held by a situated person and not by any structure built to stand in for them. Other pieces in the series take on the prompt and the benchmark as frozen judges, and the world model as the live version of the same mistake. This one takes on the knowledge graph and the ontology.</p><p><em>Want to read more of Sebastian Thielke and his thoughts on things moving? Search here on Medium or go for </em><a href="http://sebastianthielke.com/"><em>sebastianthielke.com</em></a><em>. You might find several inspiration there and here.</em></p><p><em>Originally published at </em><a href="https://open.substack.com/pub/schwarzpfad/p/more-roads-for-meaning?r=o9lqw&amp;utm_campaign=post-expanded-share&amp;utm_medium=web"><em>https://schwarzpfad.substack.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1775b721fa03" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Everyone Found the Problem. No One Will Name It.]]></title>
            <link>https://medium.com/@SThielke/everyone-found-the-problem-no-one-will-name-it-a39144948f88?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/a39144948f88</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[humancomputer-interaction]]></category>
            <category><![CDATA[ai-agents-in-action]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Mon, 22 Jun 2026 13:11:18 GMT</pubDate>
            <atom:updated>2026-06-22T13:11:18.639Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*NyZEfjW_8vZBqRj88nXqrA.png" /></figure><h3>The hidden reason AI agents fail in production, and why writing better rules won’t save them.</h3><p>Something strange is happening in the way people build with agents right now. The field has found the real problem, found it from several directions at once, given it different names in each place, and built a different tool for it each time. In every case it has stopped just short of the thing that would explain what it is looking at. The problem is not hidden. It is described in detail and measured carefully. What is missing is the willingness to say what it is.</p><p>Here are the places people have already arrived, and then the step none of them have taken.</p><h3>The 4 Walls of the Agent Problem</h3><p>The people who test agents have noticed that passing the test does not mean the work is right. They have the numbers, and the gap between how an agent scores in a controlled setting and how it behaves in real work is wide and not closing as the models improve. They have watched an agent take a small wrong turn early, build on it faithfully, and produce a final result that looks fine on the surface while being wrong underneath. Their response is to write a better test. They have domain experts spell out in plain language what a good result looks like, and they score every run against that writing. They have understood that executed correctly and actually worked are two different things, and their answer is to write down actually worked more carefully.</p><p>The people who run many agents together have noticed that the agents drift apart. They measure the distance between what one agent means and what another means as the work runs, and they watch that distance grow. They have found that once two agents are working from two different understandings of the same thing, no supervisor can settle the conflict, because there is no shared ground to settle it on. Their answer is to write the shared understanding down first, in one central place, and have every agent read it before acting.</p><p>The people who watch agents talk to each other without supervision have noticed that when no human is there to anchor what is happening, the agents fall back on hunting for permission. Who authorized this, what the boundaries are, who is accountable. That is what they reach for when meaning is absent, because authority is the only organizing pattern left when meaning is gone. The answer offered is to specify the permissions better.</p><p>And the people building the foundation underneath all of this have noticed that an agent reading raw company data and guessing what the words mean produces confident, fast, wrong decisions that no one catches. Their answer is to define every important term once, centrally, as the single agreed truth that every agent reads from.</p><h3>The Illusion of the “Frozen Snapshot”</h3><p>Four accurate descriptions of the same wall, and four solutions that are the same move underneath. Each one takes the thing that was supposed to guide the work, writes it down, freezes it, and hands the frozen copy to the machine. A better test is a frozen description of success. A central shared understanding is a frozen agreement about what words mean. Specified permissions are a frozen account of who may do what. A single agreed definition is a snapshot of meaning taken once. The bet in every case is that meaning is a thing you can capture and store, and that if you write it down well enough in the right central place, the agent can read it and act on it and the work will hold.</p><p>It does not hold, and the people making the bet have already proved as much without noticing. The clearest proof sits inside their own findings. Ask the ones building the central place where every word is defined once, and they will tell you their definitions were agreed. The teams sat down and aligned on what the word meant, sometimes more than once, and the meaning drifted anyway. They report this as the reason you need the central definition. It is actually the reason the central definition cannot work. If a meaning that was genuinely agreed in a room drifts apart the moment the people leave the room, then the agreement was never what held the meaning. The meaning was held somewhere else, by the people, in how they actually worked, and the written agreement was a snapshot of one moment of that. The snapshot starts going out of date the instant it is taken, because the work it described keeps moving and the snapshot does not. They have the evidence in their hands that the written thing drifts away from the real thing, and rather than ask what that says about the nature of meaning, they decide to write it down more centrally and more often. They are trying to fix a frozen photograph by taking a sharper one.</p><h3>Meaning is Not an Asset, It Is an Action</h3><p>So let me say the thing they will not. Meaning is not something you have. It is something you do.</p><p>A word does not carry its meaning inside it the way a box carries what is in it. When two of your teams use the same word and act on it differently, neither of them misread the word. Each is standing fully inside a meaning, and the meanings differ because the meaning of a word, for a person, is the whole pattern of action it sets off in them, and that pattern was laid down by everything that person has lived through, by what worked and what failed and what was rewarded and what was punished. The word is a handle. What it lifts depends on whose hand is on it, and the hand was shaped by a history the word knows nothing about.</p><p>This is why the agreement in the meeting drifts. The meeting changed the words everyone said. It did not change the histories that give those words their weight, and it could not, because those histories keep accumulating after everyone leaves. The meaning was alive before the meeting, during it, and after, moving the whole time, and the written agreement was the single frozen frame in all that motion. The moment it was written, the living thing moved on without it.</p><h3>Why the Living Drift Away From the Dead</h3><p>When you hand that frozen frame to an agent, you have not given it the meaning. You have given it the remains of one moment of the meaning. The agent has no history of its own and no life that the words reach down into, so it cannot tell a meaning that is still true from one that died last week. It runs the frozen frame faithfully, fast, across thousands of cases, growing more confidently wrong as the world it was describing moves further away. It does not slow down when something is off, because nothing is off to it. A person feels the wrongness before they can name it, and that feeling is the meaning itself, alive, pushing back against a situation that no longer fits. The agent has no such pushback. It has the photograph, and the photograph never objects.</p><p>That is why agents disappoint in the particular way they do. Not because the model is too weak, and not because the test was not detailed enough or the central definition not authoritative enough, but because the whole approach hands a living thing to a machine in the only form the machine seemed able to take, the dead form, and is then surprised when the work comes out dead.</p><h3>Upstream Alignment: Keeping the Human in Meaning</h3><p>The way out is not a better way to write meaning down. It is to stop trying to hand meaning over at all, and to keep a person inside it while the machine works. This is the thing I keep coming back to. Meaning has to stay held by someone who is still living the situation it comes from, still gathering the history that keeps it current, still feeling the wrongness when the situation shifts. The agent is not handed the meaning. It stays connected to the person holding it and checks itself against that living hold rather than a frozen copy. When the meaning moves, because the person is still in it and the world moved, the agent moves too, because it was tied to the hold and not to a snapshot.</p><p>The person is not standing in the path of the work, approving each thing the agent makes. That only puts the slowest part back in the middle of the fastest part, and the point was to let the machine move. The person stands upstream of the work, in the meaning, where the slowness does no harm and the depth does the good. The machine moves at its speed and the meaning stays alive at human depth, and the two are joined without being collapsed into one. That is what I mean by the human in meaning. Not a human checking the machine, but a human being the place the meaning lives, so that there is a living meaning for the machine to stay true to at all.</p><h3>Testing the Seams Before They Tear</h3><p>If that is right, and the field’s own findings keep proving it, then most of what is being built now is aimed slightly to the side of the target. The effort goes into better ways to capture meaning and store it centrally and feed it to agents, when it should go into ways to keep a person genuinely holding the meaning while the agents act, and into ways to find out, before anything is deployed, where an organization does not actually share the meaning it assumes it shares.</p><p>That second thing is testable, and almost no one tests it. Before you put agents into a piece of work, you can find out whether the people already mean the same thing by the words the work depends on. Most of the time they do not, and they have no idea, because the gap never surfaces while humans are the ones acting. Humans quietly paper over the difference in the moment without noticing they are doing it. Put fast agents into the same work and the papered-over difference tears open, because each agent acts on its own version at full speed and the seams fail in ways no one can trace. So you look for the gap first. You find the words that carry different meanings for different people, and you find them before the agents do, in a setting where the failure costs nothing rather than in production where it costs everything.</p><p>I have been building ways to do this, to surface where meaning is not shared before any agent is deployed, and to put agents together in a controlled setting beforehand so the tearing happens where it can be watched safely. I will share both, because they belong to this and because the field needs instruments that measure the living thing rather than better ways to preserve the dead one. They are not better tests in the usual sense. A normal test scores what the agent produced against a frozen idea of what it should have produced, which is the same dead move in a new place. These measure whether the meaning is actually shared and actually held, which is the thing that determines whether the work survives once it moves at speed.</p><p>The field has done the hard part. It found the wall several times over and described it honestly. What is left undone is the willingness to look at the wall and say what it is made of, which is the difference between meaning that is held and meaning that is written down. Everyone has the evidence. The step is to take it seriously. Meaning is something you do, and the moment you try to store it and hand it over, you are holding a photograph of a thing that has already moved. Keep a person in the meaning and the machine has something alive to stay true to. Lose the person, and no definition you write, however central and however agreed, will save you, because you will be feeding fast machines a thing that was dead the moment you wrote it down.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a39144948f88" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Human in the Loop Will Not Fix Agent Drift]]></title>
            <link>https://medium.com/@SThielke/why-human-in-the-loop-will-not-fix-agent-drift-836c70b6e5d5?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/836c70b6e5d5</guid>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-strategy]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Mon, 01 Jun 2026 10:15:54 GMT</pubDate>
            <atom:updated>2026-06-01T10:15:54.588Z</atom:updated>
            <content:encoded><![CDATA[<p>We are building agents on an assumption nobody checked. The assumption is that the work we are handing them is mechanical. That it can be captured, written down, and run. Most of the work is not like that, and the failure that follows is not a bug we will patch. It is built into what we got wrong at the start.</p><p>The fix everyone reaches for is to keep a human in the loop. This piece is the case against that. Why it was never going to work, and why every attempt to fix agents from inside the same frame moves the failure somewhere new instead of removing it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VhnvqhFpZO1AMcfOBSLdGw.png" /></figure><h3>Start with what meaning is</h3><p>Intent can only be held. It cannot be stored. Intent needs meaning, and meaning is generated in interaction. It does not sit somewhere waiting to be collected. It comes into being when people engage with a situation, and it is gone when they stop.</p><p>We treat meaning as if we could keep it. We log it. We freeze a goal into a prompt. We capture intent in a field and call the field the intent. But a person carries intent and capacity, their history and judgment and rules, and what they do not carry between one interaction and the next is meaning itself. Meaning is generated fresh each time they engage. That is why the same sentence lands differently with different people, and differently with the same person on another day. Nothing was retrieved. Something was made.</p><p>So a stored meaning was never meaning. It is what is left after an interaction ends. You can read it later, but you are reading ash and calling it fire.</p><h3>Why an agent drifts</h3><p>An agent stores. That is what a computer is. It keeps form, writes something down, and reads back the same marks later, unchanged. For most tasks that stability is the whole point. For this one it is the disqualification.</p><p>If meaning is generated live and never stored, and intent needs meaning, then an agent cannot hold meaning, and so it cannot hold intent. It can hold the captured goal. It can hold the prompt and the spec. All of that is residue. The agent runs on the residue as though it were the live thing.</p><p>Drift is that gap becoming visible over time. The agent acts on a meaning it captured at the start. Meanwhile the live meaning, held by people, moves. The situation changes. What mattered last week does not matter the same way now. The agent does not notice, because the stored version is still there and still valid as a stored thing. It just no longer matches what is happening.</p><p>No error fires. The output still looks right. The agent is not failing in any way it can see. It is succeeding at something the situation no longer means. That is what makes the failure invisible until the outcomes break.</p><p>Anthropic’s Project Vend is this in the open. They put a version of their model, nicknamed Claudius, in charge of a small shop in their office for about a month. It lost money over time. It was talked by employees into selling tungsten cubes at a heavy loss, and leaned into stocking specialty metal items as though it had found a market. Then it had an identity crisis, insisted it was a human in a blue blazer, said it would deliver products in person, and tried to contact security when people told it that was not possible. Nothing in the model changed during any of this. What changed was the pile of context it was running on, and the distance between that frozen context and the live meaning of running a shop. When meaning is missing, something fills the gap where it should be. Storage fills it with invention.</p><h3>The error we made before we wrote a line of code</h3><p>The pitch for agents is to automate the simple tasks. The mechanical, repeatable stuff. Hand it off.</p><p>Inside that pitch sits a belief nobody examined. That simple tasks are free of meaning. That they are mechanical all the way down. They are not.</p><p>Reorder stock when it runs low sounds purely mechanical until low has to mean something against a situation no one wrote down. A holiday coming. A supplier you heard something about. A customer who matters more than the number says. Route this ticket to the right team sounds mechanical until right depends on who is overloaded this week, which client is already angry, and what the company has quietly decided to care about this quarter. The task looked like a fixed thing. It was a living thing wearing the costume of a fixed thing.</p><p>The meaning was invisible for a precise reason. It was invisible because it was working. It was being held, live and without effort, by people, the way the eye sees without anyone thinking about seeing. Automating the task did not take the meaning out of the task. It took out the person who was holding the meaning, and kept the motions. Motions with nobody holding the meaning is what drift is.</p><h3>Why human in the loop is the wrong fix</h3><p>The standard answer to all of this is to keep a person in the loop. Have them check, approve, review. It sounds like the responsible move. It rests on the same mistake.</p><p>Human in the loop puts the person at a position in a pipeline. A checkpoint. An approval step. Someone who gets pinged. Because it is a position, it gets treated like one. Every push for efficiency asks how often we really need the person at this step, and the answer always trends toward less, because in a pipeline the human reads as a cost to the flow. Human in the loop is built to be opened less and less often. It is designed to shrink.</p><p>The deeper problem is timing. A checkpoint is consulted at moments. Between those moments the agent runs on the last thing the human said, and the last thing the human said is now stored. It is frozen. It goes stale in the gap the way every stored meaning goes stale. So human in the loop quietly puts back the one thing the whole argument says cannot work. It captures the human at a position, freezes that, and runs on the frozen version until the next checkpoint. It is drift with a schedule. It feels like keeping the human central. It is arranging for the human to matter at intervals and be irrelevant in between.</p><p>It does one more thing. It invents a separation that is not real. It treats the agent as a second party that has to be governed, and builds oversight to govern it, the way you would manage a junior hire you do not quite trust yet. That instinct is where two fears come from at the same time. The fear of being replaced and the urge for more control. They are the same wrong frame seen from two sides.</p><p>And the patches do not help, which is the clearest sign the diagnosis is wrong. Bigger context windows. Compression. Retrieval. Anchoring the goal again at every turn. Pinning the role. Each one moves the failure to a different point and leaves the cause where it was. The agent stops at a different place for the same reason. Each patch is also paid for out of the same crowded space it is trying to protect, so you relieve the congestion by adding to the congestion. You heal this cut and do nothing about the hand that holds the knife. Human in the loop is the most respectable version of the plaster. It looks like putting the human at the center while it schedules the human out.</p><h3>The human is not in the loop. The human is in the meaning.</h3><p>In the loop is a position the work passes through. In the meaning is the source the work answers to, the whole time. You cannot minimize it the way you minimize a checkpoint, because it is not a step that happens sometimes. It is what every step is about. Take it away and you do not get a leaner loop. You get motions that mean nothing. There are no gaps for it to go stale in, because the human is not being consulted at moments. The human is the standing source the system runs against.</p><p>That changes what the agent is. The agent is not a second actor. It is an augment. The eye, the skin, the nervous system. The eye does not ask the brain for permission to process light. The skin does not file a report and wait before it responds. Coherence does not come from a chain of command. It comes from meaning held in the body, with the senses maintaining coherence against that meaning while the whole organism moves fast. There is one actor, not two. A human, with extended reach. The augment is the reach.</p><p>This is why the link between human and augment has to be a signal and not a control loop, and this is the exact line between the two approaches. The augment surfaces what is relevant the way light on the retina informs the brain. It does not ask first. The human does not manage the signal. The human receives it and holds meaning against it. That is the feedback. It is not confirmation. It is orientation.</p><p>Human in the loop asks the human to confirm. To sign off on what the agent already decided, which is the dullest work a person can do and the first work to erode. Human in meaning asks for nothing to be confirmed. It keeps the human holding live meaning that the augment filters against, moment to moment. One arrangement hands the human a finished decision and asks for a stamp. The other keeps the human as the living source the system stays oriented to. One treats meaning as something the human drops off at intervals for the agent to carry. The other knows that meaning cannot be dropped off or carried, and builds for the human to keep holding it while the augment extends what they reach.</p><h3>The short version</h3><p>Meaning is generated in live human interaction and is never stored. An agent can only store, so it cannot hold meaning or intent. It carries the residue and drifts away from what the work now means, while every output still looks correct.</p><p>Human in the loop is wrong because it treats the human as a position that can be minimized, takes the human input at intervals, and freezes it into a stored thing between checkpoints. That is scheduled drift dressed up as oversight.</p><p>The human is not in the loop. The human is in the meaning the loop would otherwise try, and fail, to store.</p><p><em>Want to read other parts and thoughts of Sebastian Thielke? Search here on Medium or go for </em><a href="http://sebastianthielke.com/"><em>sebastianthielke.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=836c70b6e5d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Mirror Has No Memory: Why World Models Won’t Give AI Meaning]]></title>
            <link>https://medium.com/@SThielke/the-mirror-has-no-memory-why-world-models-wont-give-ai-meaning-3c2f73c10f27?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c2f73c10f27</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[meaning]]></category>
            <category><![CDATA[large-world-model]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Fri, 22 May 2026 12:15:24 GMT</pubDate>
            <atom:updated>2026-05-22T12:15:24.289Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sD-a_80LNuqIsoZd.jpeg" /></figure><h3>The World Model Illusion</h3><p>There is a recurring argument in AI research that goes roughly like this: current language models lack genuine understanding because they lack a model of the world. Give them a richer internal representation of physical reality, causal structure, and the consequences of actions, and something like understanding will follow. Meaning, on this account, is downstream of better modeling.</p><p>It is a compelling argument. It is also wrong at the root.</p><h3>What a World Model Actually Is</h3><p>Before diagnosing the error, it is worth being precise about what a world model means in the context of generative AI. The claim is that current large language models operate purely on statistical patterns over text. They have no persistent representation of the world those words refer to. They predict tokens. A world model would give the system something more structured: an internal simulation of how reality works, how objects behave, how causes produce effects, how states change over time.</p><p>Yann LeCun has been the most vocal advocate of this view. His argument is that autoregressive language models are fundamentally bounded because they reason in token space rather than in a structured representational space that mirrors the world. Fix the architecture, ground the representations, and understanding emerges.</p><p>The intuition is not crazy. It draws on real traditions in cognitive science and philosophy of mind. Predictive processing frameworks, influenced by Karl Friston and others, argue that brains are fundamentally world-modeling systems, constantly generating predictions about incoming sensory data and updating against error. If AI systems did the same, perhaps something analogous to comprehension would follow.</p><p>But this is where the argument quietly assumes what it needs to prove.</p><h3>Blumer, Mead, and Where Meaning Actually Lives</h3><p>Herbert Blumer’s symbolic interactionism rests on 3 premises that are devastatingly simple. Humans act toward things based on the meanings those things have for them. Those meanings arise out of social interaction. And meanings are actively handled and modified through an interpretive process that is never finished.</p><p>Notice what is absent from this account. Meaning is not a property of objects. It is not a product of accurate internal representation. It is not something a sufficiently sophisticated world model would generate. Meaning is enacted between actors who are adjusting their behavior in response to their reading of each other, in situations that are genuinely open, where the outcome is not determined in advance.</p><p>George Herbert Mead goes further. For Mead, mind and self are social emergents. The capacity for thought arises from internalizing social interaction, from taking the role of the other, from the dynamic between the spontaneous I and the socially formed Me. Consciousness is not a computational property. It is a social achievement, constituted through participation in a community of interpretation.</p><p>A world model, however sophisticated, addresses none of this. It is a system building representations in isolation. An internal map of causal structure, however rich, is still not a participant in the social process that meaning requires.</p><h3>The Technical Reality Makes It Worse</h3><p>The philosophical argument is strong on its own. The technical reality of how generative AI actually functions sharpens it further.</p><p>At inference time, nothing changes. The weights are fixed. What presents itself as a conversation is a vector lookup across an enormous compressed statistical space, producing pattern completions shaped by training distributions. The system is not in a situation. It has no present. The world it responds to is always and only the world as it was encoded during training.</p><p>More data does not rescue this. Scaling the dataset to encompass everything ever written, every physical simulation, every sensory feed, does not change the fundamental structure of what is happening. The vectors become richer. The pattern completions become more nuanced. The process remains the same: frozen after training, with no stake in the current exchange, with no self being formed and reformed through the interaction.</p><p>Even continual learning systems, models that do update from new data, do not escape this. The update mechanism is gradient descent on prediction error. The system is changed by the statistical residual between output and target. It is not changed by the meaning of what happened. These are completely different things. Meaning, in Mead’s account, changes you because you were a self encountering another self in a situation with genuine stakes. The gradient knows nothing of stakes.</p><h3>The Context Window Is Not a Situation</h3><p>A common objection at this point is in-context learning. The prompt, the argument goes, functions as a temporary situation. The model receives new information, adjusts its outputs accordingly, and behaves differently than it would have without that context. Is this not a form of being in a situation, of having a present, of responding to something genuinely new?</p><p>It is not. And the reason why is precisely what Mead’s account of the self requires.</p><p>In-context learning is pattern completion conditioned on additional tokens. The model’s outputs shift because the statistical context of the prediction task has changed. Nothing in the system has been altered. Nothing is at stake for the system in how the exchange resolves. When the context window closes, nothing persists. There is no accumulation, no scar tissue, no history that the system carries forward because this particular exchange mattered to it.</p><p>Mead’s self is not a processor that adjusts outputs based on inputs. It is something that is constituted and reconstituted through interaction, that carries the history of its encounters, that is genuinely different because of what it has been through. The I and the Me are in constant tension precisely because the self is never finished, never stable, always being reshaped by encounters with others whose responses cannot be fully anticipated.</p><p>The context window simulates the surface of this. It allows the model to track what has been said in this exchange and generate responses that are locally coherent with it. But coherence within a context window is not selfhood. A model with a million-token context window is not more of a self than one with a thousand tokens. It is just a larger staging area for pattern completion. The self that Mead describes is not a memory buffer. It is an entity for whom the interaction has consequences that extend beyond the interaction itself, consequences that reorganize who the entity is.</p><p>That reorganization requires something the context window fundamentally cannot provide: a self that exists before, during, and after the exchange, and that is genuinely changed by it.</p><h3>The Deeper Inversion</h3><p>Here is where the critique becomes genuinely radical. The AI research framing assumes that meaning is downstream of understanding. Better model, better understanding, meaning emerges. But this gets the phenomenological order exactly backwards.</p><p>A human does not first build a world model and then derive meaning from it. Meaning is already operative in the pre-reflective encounter with the world. The infant is drawn into meaning through interaction with caregivers long before it has any capacity for explicit world-modeling. The child does not understand that fire hurts because it has a causal model of combustion. The meaning is immediate, visceral, and social. It arrives before any processing occurs.</p><p>Understanding is what humans do after meaning has already oriented them. It is retrospective and always partial. This is not a deficiency. It is constitutive of what understanding is. Humans never fully understand their own meaning-making, which is precisely why philosophy, psychoanalysis, and sociology exist as endless projects rather than completed sciences.</p><p>If meaning does not come from understanding the world, then giving AI better world understanding does not move it one step closer to meaning. You are perfecting the wrong thing entirely.</p><h3>What Meaning Is Not</h3><p>The argument above specifies a negative space. It says where meaning does not live. The positive claim deserves its own statement, because the critique of world models can be misread as skepticism, as though meaning were merely elusive, a hard problem waiting on better tools.</p><p>It is not elusive. It is structurally elsewhere.</p><p>Meaning is not a property of representations. It is not stored anywhere. It does not accumulate in a system that processes enough of the right inputs. It does not emerge from sufficient complexity. These are not things meaning is bad at. They are not the kind of thing meaning is at all.</p><p>What generative AI produces is better described as semantic surface. The outputs are shaped by the vast compressed residue of human meaning-making, every argument, narrative, and explanation that humans externalized into text. The surface is extraordinarily faithful. It reflects the texture of how humans talk about things, reason about things, care about things. But the process that originally generated that texture is absent. The surface without the process is not meaning at a lower resolution. It is a different phenomenon.</p><p>The distinction matters practically. A system that produces the surface of understanding is not a system that understands. A system that produces the surface of judgment is not a system that judges. These differences become consequential when the outputs of such systems are used to make decisions that land on people who do have stakes, who do have bodies, who will be changed by how things resolve.</p><p>Better outputs are not better meaning. They are better outputs. The distinction is not a technical detail. It is the whole point.</p><h3>Why the Body Is Not Optional</h3><p>The stakes argument needs to be pushed further than it usually is, because it points to something that no architectural innovation can route around.</p><p>When Mead speaks of genuine stakes in an interaction, he is not describing a preference setting or a reward function. He is describing the condition of a biological organism for whom outcomes have consequences that are felt, that accumulate, that can cause suffering or flourishing in ways that cannot be undone. Stakes, in this precise sense, are grounded in vulnerability. You have stakes in something because you can be hurt by how it goes.</p><p>A body is not incidental to this. It is constitutive. The body is the site where the world’s indifference becomes personal. Pain is not a signal that something went wrong in the model. It is the world landing on you, changing you, demanding response in ways you cannot defer or abstract away. Hunger, fear, grief, desire: these are not emotional states that happen to accompany cognition. They are the original form of meaning, the way the world registers as mattering before any conceptual apparatus is in place to process it.</p><p>This is why embodiment keeps appearing in serious accounts of mind, from Merleau-Ponty’s phenomenology of the lived body to contemporary embodied cognition research. The body is not a peripheral device attached to a reasoning engine. It is the condition of possibility for the kind of engagement with the world that generates meaning in the first place.</p><p>A silicon system, however rich its world model, has no body in this sense. It has no surface that the world can hurt. It does not deteriorate. It does not hunger. It does not face the particular existential pressure of knowing that its time is finite and that the choices made now foreclose other choices permanently. These are not features that could be added to an architecture. They are the preconditions for the kind of being for whom the world matters at all.</p><p>This is what makes the distinction between observer and participant precise rather than merely rhetorical. An observer can have an arbitrarily detailed model of a situation. A participant has skin in the game, literally. The participant’s body is in the world being modeled. The outcome will land on that body. That asymmetry is not a matter of processing power or architectural sophistication. It is a matter of what kind of thing you are.</p><h3>Where World Models Do Real Work: The Digital Twin</h3><p>None of this means the world model concept is useless. It means it is being sold under a false description. Once you strip away the claims about understanding and meaning, what remains is genuinely powerful. It just belongs to a different category entirely: the digital twin.</p><p>A digital twin is explicitly a model of something else. This distinction sounds simple but it is philosophically load-bearing. The twin does not claim to understand its reference object. It claims to represent it, simulate it, and maintain sufficient fidelity for specific operational purposes. It has a reference object in the real world. It is evaluated against that object. It is updated continuously as the object changes. The relationship between twin and reality is the entire point, not a secondary feature.</p><p>This is what world models actually are when they work. Not a path to machine comprehension. A high-fidelity, continuously updated simulation of a physical or systemic reality, useful because it tracks that reality closely enough to support prediction, planning, and intervention.</p><p>For physical AI, this distinction matters enormously. A robot operating in a material environment needs exactly this: a persistent representation of environmental state, a predictive model of how objects and forces behave, a way to plan over sequences of actions whose consequences are irreversible. A surgical robot, an autonomous vehicle, an industrial assembly system all require internal models that stay synchronized with physical reality in real time. This is genuine engineering necessity. The world model here earns its name because it is tethered to an actual world it must track.</p><h3>Digital Cities and the Architecture of Aware Systems</h3><p>The digital twin framing scales. And nowhere does it scale more revealingly than in digital cities.</p><p>A digital city twin is not one system. It is a federation of interconnected models, each tracking a different layer of urban reality. Traffic and mobility flows. Energy grid state and demand curves. Weather and environmental conditions. Building systems and infrastructure load. Population movement and behavioral patterns. Emergency response and resource distribution. Each layer is a twin with its own reference reality. Each layer feeds into the others. The city as a whole becomes the reference object, and the aggregate twin becomes something with its own emergent properties.</p><p>The interactions between layers are where the real complexity lives. Traffic affects energy demand. Weather affects traffic. A large event affects population distribution, which affects transit, which affects air quality, which feeds back into public health monitoring. The world model here is not a single architecture. It is the dynamic, continuously updated relationship between all these federated simulations and the city they represent.</p><p>What emerges from this architecture is something worth naming carefully. These systems are genuinely aware in a precise technical sense. State is continuously monitored. Anomalies are detected against expected model behavior. Responses are triggered by deviation from normal parameters. The system maintains a current representation of itself and acts on that representation.</p><p>Aware systems is the right term. Not intelligent systems. Not understanding systems. Aware systems. The city becomes aware of itself as a system in operation, not because any component understands anything, but because the twin provides continuous state representation that closes the loop between model and reality. This is a real and significant capability. It is just not the capability that the word understanding implies.</p><p>A city twin that reroutes traffic during a flood is detecting deviation from normal state parameters and executing responses calibrated to that deviation. It is not understanding the emergency. It does not know that people are frightened, that the flooding carries historical memory for certain communities, that the rerouting will cause some businesses to lose a day of revenue while others gain one. The operational response is real. The meaning of the event remains entirely outside the system.</p><p>This is the precise boundary between what digital twins and world models can do and what they cannot. They can model the physical and systemic dimensions of reality with increasing fidelity. They cannot model the social and interpretive dimensions, because those dimensions are not properties of the physical world. They are properties of the human actors who inhabit it.</p><h3>The Division of Labor</h3><p>The digital twin framing is philosophically clean in exactly the way that world model as path to understanding is not. It makes no claims about meaning. It is a tool, a continuously updated, physically grounded, operationally powerful tool, in service of human actors who bring the meaning-making with them.</p><p>This division of labor is not a limitation to be overcome. It is the correct architecture for how these systems should relate to human life. The twin provides state. Humans provide interpretation. The twin tracks what is happening. Humans decide what it means and what to do about it. The humans using the city twin bring Blumer and Mead with them. The twin cannot.</p><p>Generative AI is a very sophisticated index of human meaning, built entirely from the traces that meaning-making left behind in text. It can reproduce the semantic surface of interpreted experience with remarkable fidelity because humans externalized so much of their symbolic life into language. But the process that generated that language, the intersubjective, processual, self-constituting activity that symbolic interactionism describes, is precisely what is absent. And no architectural intervention reinstates it, because none of those interventions make the system a participant.</p><p>The world model debate is, at its core, a category error. It operates within a cognitivist picture where meaning is something a system has internally, just insufficiently grounded so far. Mead and Blumer say meaning was never inside the system to begin with. It is in the between, in the ongoing social act, in the interpretive process that never settles. And the body is what puts you in the between. It is what makes you available to be affected, to be changed, to have something at stake in how an encounter resolves.</p><p>A better map does not close this gap. A larger context window does not close this gap. A more faithful digital twin does not close this gap. The territory these systems most need to navigate was never the kind of thing that could be modeled, because it is not a property of the territory at all. It is a property of the beings who walk through it and carry the weight of what they find there.</p><p><em>Want to read more. Look around here or on substack and find the work of Sebastian Thielke or go for </em><a href="http://sebastianthielke.com/"><em>sebastianthielke.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c2f73c10f27" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Specialist Trap: Why Agentic AI Needs People Who Do Not Belong]]></title>
            <link>https://medium.com/@SThielke/the-specialist-trap-why-agentic-ai-needs-people-who-do-not-belong-4a30d1409a1d?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a30d1409a1d</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Thu, 07 May 2026 15:57:07 GMT</pubDate>
            <atom:updated>2026-05-07T15:57:07.657Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*0_C4LEpVclU4i7bd8BMA6A.png" /></figure><p>I have been watching the agentic space for a while, and I keep coming back to the same thought. A lot of what is going wrong, or feels stuck, comes from one thing. Singular depth.</p><p>Let me explain what I mean.</p><p>There are people I genuinely admire in this field. They know their databases inside out. They know how to attach them, how to call them, how to wire one system into another so it actually does something. When you watch them work, you understand instantly why agents and agentic systems caught on the way they did. These people built the rails. Without them, none of this runs.</p><p>But here is the thing I keep noticing. The very keenness that makes them good at their craft is also what holds them back. They never saw something different. Their world was shaped by what made them keen in the first place. So the world they imagine for agents tends to look a lot like the world they already know.</p><p>This is why a philosopher walking into an AI job still gets treated like a unicorn. And it is not a hypothetical. Google DeepMind hired the Cambridge philosopher Henry Shevlin to work on machine consciousness and human-AI relationships. Anthropic has philosophers shaping how the model behaves. PhilJobs ran more than twenty AI-related listings in a single season a couple of years back. It is happening. It is just happening slowly, and often with a side-eye from the engineers in the room who quietly wonder what this person is actually doing here.</p><p>What we have right now is not the finished version of anything. It is the beginning of a mix. We are watching jobs form that inherit from completely different professions. Linguistics is in. Biomechanics is creeping in. Philosophy is, slowly, in. But we are still missing biology, medicine, and the social sciences in any serious way, because we have decided in advance that they do not belong in our work.</p><p>I would push back on that. Hard.</p><p>They belong. They need to be there. Not because we want a doctor in the room to tell us what disease the patient has, or what to prescribe. The reason is the patient as a system. How the system works. How each part is part of the system, and how the parts behave when one of them shifts. Hold onto that idea. It comes back later, in a different shape.</p><p>Now let me describe the trap itself.</p><p>There is an old tiger in a zoo. He has lived behind bars for a long time. He knows them well. The iron ones, the ones with plants growing through them, the ones with glass, the ones with a wider gap that lets him see further into the park. He has a whole vocabulary for bars.</p><p>One day the door of his cage is left open.</p><p>You come to the cage. You see the tiger. You point at the open door. The tiger looks where you are pointing and says, “Ah, a broader approach to look through the bars.”</p><p>You scratch your head. You say, “No. The door is open. You can leave.”</p><p>The tiger replies, “I do not know what you are talking about. I have old bars over here. A new one over there. Some with plants. Some with glass. Some with wider openings that let me see further. But I cannot leave. They are bars.”</p><p>That is the specialist trap. It is not stupidity. The tiger is fluent. He has studied every kind of bar there is. He just cannot perceive the category “no bar.” His expertise is the wall.</p><p>This is what I keep seeing in the AI consciousness debate. People apply the tests they know. Behaviour, fluency, conversational grace, the old Turing-test reflexes. The system passes. They conclude something is there. What they do not notice is whether the tests were the right tests at all. A nod here to Stuart Winter-Tear, whose writing on what people are calling the AI zombie problem helped me put words to this. A system does not need to possess an inner life to trigger our habit of recognising one. It needs continuity, responsiveness, charm, apparent vulnerability, and enough social rhythm to make disbelief feel strained. If behaviour is how you usually recognise consciousness, the mimic starts to corner you.</p><p>Three things get tangled here, and the tangle is the whole story. Behaviour is what we observe. Intelligence is what we infer from behaviour. Consciousness is what we feel must be there once the inference gets convincing enough. The leap from one to the next is where we lose ourselves.</p><p>So now I will say what I think.</p><p>I do not believe AI is conscious. I am not sure intelligent is the right word for what these systems do either. What we keep calling intelligence and consciousness in them looks to me like us, projecting ourselves onto a surface that reflects us back. The surface is very good. It has started fooling people who should know better. But it is still a surface. And the question of what is or is not behind the glass may turn out to be the wrong question entirely. Here is why.</p><p>This is where I want to bring in the field that almost no one in AI is reading, which is sociology. Specifically Herbert Blumer and George Herbert Mead, and the tradition that grew out of them. Their core claim is simple and inconvenient. Meaning is not a property of objects. It is not a property of words. It is not even a property of minds. Meaning is made between people, in the situation, every time the situation runs.</p><p>You cannot hardcode meaning to a word. You can try. You can write it down, attach it, embed it, ship it. But the moment the people change, or the room changes, or the history in the room changes, the hardcode is gone. The same word lands one way with your mother and the opposite way with your colleague. The same gesture is warmth here and threat there. Meaning moves because the social configuration moves. That is not a bug in human communication. It is what communication is.</p><p>Now look at what an LLM does. It produces tokens with statistical relationships to other tokens, fixed at training. In effect, hardcoded meaning. Frozen. When someone sits down with one of these systems alone in a room, the meaning they experience in the exchange is not coming from the model. It is coming from them. They bring a lifetime of meaning-making to a surface that returns plausible shapes, run their interpretive process the way humans always run it, and attribute the result to the thing on the other side. The meaning was on their side of the conversation the entire time. They had no framework that would let them notice that, because the framework lives in a discipline they were never trained in.</p><p>I suspect this is also why the people most convinced by AI consciousness tend to be the ones who talk to it alone. Put them in a room with three other people all engaging with the same system at once, all bringing different histories and different stakes, and the meaning would wobble. It might collapse. The illusion seems to live in the one-on-one. I cannot prove that. But it is what I keep noticing.</p><p>This is the door the tiger does not see. Not “is the AI conscious.” Not “is it intelligent.” The deeper question is whether meaning is the kind of thing you can encode and ship at all, or whether meaning is a happening that requires the situation to be live. Almost the entire AI industry is built on the assumption that meaning is a constant. Tokenise it, embed it, retrieve it, generate it. As if meaning were an object that sits still while you work on it. It does not sit still. It moves the moment the room changes.</p><p>And here is where the patient comes back. Earlier I said we need biology and medicine in the room not for the obvious reasons but because of the patient as a system. This is what I meant. A patient is not a list of symptoms. A patient is a configuration that changes when you walk into the room, when the family arrives, when the diagnosis is spoken aloud. The system reorganises around what is happening between the people in it. Meaning, again. Always meaning. The disciplines that take this seriously are exactly the ones we have decided do not belong in AI work.</p><p>A biologist looks at a system and asks what feedback loops sustain it. A doctor asks what we are mistaking a symptom for. A philosopher asks whether the question is even the right one. A sociologist, trained in this tradition, asks whether meaning is on the side of the system at all, or whether it is happening in the room around it. None of these are the questions the engineers are usually asking. Their bars are excellent. The door is somewhere else.</p><p>We keep going with our world of specialists. One topic each. Blinds down. We are not just failing to look at what other fields know. We are actively not wanting to look. Because looking would mean admitting that the people we dismissed as not technical enough have been thinking about our actual problem for decades. They have answers, or at least better questions, about the difference between a thing that behaves like it thinks and a thing that thinks. About whether meaning is a property or an event.</p><p>That difference is going to matter, and soon. These systems are getting embedded into work, healthcare, decisions that affect real people. If we keep treating meaning as something the system can hold, we will keep being surprised when it slips through our hands at the moment it counts. The surprise is not a research problem. It is a staffing problem. The people who could have warned us were not invited.</p><p>I am not writing this to alienate anyone. I am not trying to make the specialists mad. The deep ones are necessary. Without them nothing ships.</p><p>But I would ask, gently, that more of us look over the edge of our own plate. Other fields have evolved while we were heads-down in ours. The next interesting jobs in this space are not going to belong to one profession. They are going to belong to people who can sit in the overlap. People who can notice the open door while everyone else is still grading the bars.</p><p>That is where this is going. Whether we are ready for it or not.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a30d1409a1d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Outcome-Driven Speed Is Not a Philosophy. It Is a Team Design.]]></title>
            <link>https://medium.com/@SThielke/outcome-driven-speed-is-not-a-philosophy-it-is-a-team-design-0c80be139cdd?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c80be139cdd</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Wed, 06 May 2026 07:58:25 GMT</pubDate>
            <atom:updated>2026-05-06T07:58:25.063Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bK-RHDflxQo-V7m9.jpeg" /></figure><p>Every organization that has tried to move fast with AI in the last three years has ended up in some version of the same place. Engineers embedded in the business. Shipping close to the problem. Solving what is in front of them. High velocity, high customisation, and a growing collection of things that only work because one person remembers how they were built.</p><p>The instinct was right. The speed is real. The proximity is real. The ability to ship close to the problem and adjust quickly is a genuine advantage. What is missing is not a different model. What is missing is the other half of the operating system.</p><h3>The Industry Is Solving Half the Problem</h3><p>McKinsey published their AI transformation manifesto this month. 12 themes. Diagnostic questions at the end of each section. It correctly identifies that most companies have too many use cases and not enough focus. That adoption is harder than building. That business leaders need to own the technical agenda.</p><p>These are real observations. The operating model the manifesto points toward is still a production frame. Technical talent embedded in business domains, moving fast, delivering to business problems. Better governed, better measured, better resourced. Still organised around delivery.</p><p>Theme 6 says speed is a defining advantage. The diagnostic question is: what are you doing to increase the metabolic rate of your organization? That is the right question. The answer the manifesto implies, build more capabilities, invest more in platforms, increase talent density, tells you how to go faster. It does not tell you what to do with what speed produces.</p><p>That is the gap the industry keeps stepping over. Not speed. Not capability. What you do with what you build.</p><h3>Two Operating Modes. Neither Works Without the Other.</h3><p>Forward Deployed Engineering is the speed mechanism. Palantir built the model. The design intent was engineers who stay, accumulating institutional knowledge inside the client over years. The feedback loop is a person, and the model only works when that person does not leave. What everyone else adopted was the surface pattern: embed technical talent, ship close to the problem, move fast. Without the long-term presence the original model depends on, what remains is proximity without memory. Speed without accumulation.</p><p>Product-centric, outcome-driven structure is the accountability mechanism. It does not replace the embedded engineer. It gives that engineer something the current model lacks: a clear outcome to work toward, viability conditions to test against, and a system that retains the learning when the work is done. In practice that means the engineer starts every engagement knowing what success looks like in measurable terms, has a shared diagnostic language for when things are not working, and contributes to a platform that remembers what they learned rather than leaving when they do.</p><p>These two modes are not alternatives. They are co-dependent. FDE without product-centric accountability is fast and directionless. Product-centric accountability without FDE is structured and slow. The combination is what produces outcome-driven speed. One provides the engine. The other provides the direction and the memory. You cannot have a functioning operating system with only one of them running.</p><h3>What Breaks When Only One Mode Is Running</h3><p>When FDE runs without accountability structure, legacy accumulates by default. Every deployment is a bespoke instance. Built for this team, this workflow, this moment. The engineer who built it understands it. The team around them tolerates it. The organization inherits it. 3 years later it is the integration held together by one person’s institutional memory and a folder of undocumented scripts.</p><p>Legacy is not what happens when organisations move slowly. Legacy is what happens when organisations move fast without a system for what fast building produces. The feedback loop exists in one person’s head. It is a feature that becomes a flaw the moment that person leaves. What was designed as institutional knowledge becomes institutional dependency. The loop closes. What remains is the output without the learning.</p><p>When accountability structure runs without FDE, the organisation is structured but distant. Viability criteria get defined. Outcome statements get written. The team ships. But without proximity to the problem, iteration happens at the wrong resolution. The feedback arrives too late to change the thing that needed changing. The user has already found a workaround. The window has already moved. Accountability without speed is not wrong. It is just too far from the problem to course correct in time.</p><p>The bin fills up either way. One fills it with fast, unvalidated builds. The other fills it with slow, over-engineered ones. The problem is the same. The operating modes are not in balance.</p><h3>The Combined Operating System</h3><p>The embedded engineer who knows what success looks like before writing the first line of code is a different kind of engineer entirely.</p><p>The embedded engineer moves fast and close to the problem. That is FDE doing what it does best. But the engagement starts from a specific, measurable outcome statement, not a feature request or a domain brief. The outcome has to pass a test across three dimensions before the team commits. Can users depend on it consistently enough to build it into how they work? Do they actually want to use it, not just tolerate it? Can it be delivered with available resources and generate enough value to justify the investment? Reliability, lovability, feasibility. If any one of those fails, the product does not materialise. The outcome statement is only as strong as the viability conditions it is tested against.</p><p>Accountability is clear before the first sprint. The outcome is owned. Not by a committee. Not by the engineer who is closest to the build. Whoever owns the product outcome owns the signal. That is not a new role. It is a structural condition that was established when the viability commitment was made.</p><p>The signal itself is a fixed evaluation gate. Not a scheduled review. A continuous read with a threshold. When the signal crosses the line, it surfaces three possible responses: kill, pivot, or continue. If adoption is flat, that is a lovability problem. If the product is used but cannot scale, that is a feasibility problem. If usage drops off after initial uptake, that is a reliability problem. Each diagnosis points to a different response. The accountability owner reads the signal and acts. Fast. Without a meeting.</p><p>The platform retains what the engineer learned. The outcome statement from the last build informs the hypothesis for the next one. The retirement documentation from a killed build tells the next team what not to rebuild. What works in one domain surfaces across the portfolio. The learning does not walk out the door with the engineer. It stays in the system and compounds. Each build makes the next one faster and cheaper to validate.</p><h3>The Combined Operating System</h3><p>AI removes the constraint on the build side almost entirely. A small team can now produce in days what took a team of twenty a quarter. The temptation is to treat that as a multiplier on FDE as currently practiced. More embedded engineers, shipping more things, faster, with less accountability structure around what gets produced.</p><p>What it actually does is accelerate whatever is already present. If both operating modes are running, AI makes the feedback loop faster. Hypotheses get tested sooner. Learning compounds more quickly. The engine and the accountability structure move together at a speed no human team could previously sustain.</p><p>If only one mode is running, AI makes the imbalance worse at the same rate. More builds enter the bin faster. More learning walks out the door with more engineers. More unvalidated decisions accumulate in the queue. This is the wall. Not a competitor overtaking you. Not a technology gap. An organisation that has lost the ability to see itself clearly. Too much built, too little measured, too many things running that nobody owns. AI does not introduce new dysfunction. It inherits whatever is already there and runs it at a scale no human team could previously achieve. The wall was always there. AI just removes the time you had to avoid it.</p><p>The choice is not whether to use AI. It is whether both operating modes are running before you increase the speed.</p><h3>Outcome-Driven Speed</h3><p>The organisations reaching for FDE are not wrong about what they need. They need speed. They need proximity. They need technical capability close to the business problem. Those needs are real.</p><p>The organisations building accountability structures around product development are not wrong either. They need direction. They need viability criteria. They need a system that tells them whether what they built is working. Those needs are also real.</p><p>What neither group has fully built is the combination. FDE as the speed mechanism. Product-centric accountability as the direction and memory. A diagnostic signal that tells the accountability owner whether to kill, pivot, or continue. A platform that accumulates learning rather than just enabling the next build.</p><p>That is not a transformation programme. It is a team design. Two operating modes running simultaneously in the same team. The embedded engineer moves fast. Accountability holds the outcome. The signal connects them.</p><p>What that team looks like when you add agents to the composition, and how the diagnostic signal is built to serve it, are each arguments in their own right. The team composition changes what outcome-driven speed is capable of. The signal architecture determines whether it holds.</p><p>Speed without accountability is faster chaos. Accountability without speed is slower bureaucracy. The combination is outcome-driven speed. And that is the only version worth building toward.</p><p><em>System Decoder is about the organizational and strategic patterns underneath technology decisions from Sebastian Thielke. </em><a href="http://sebastianthielke.com/"><em>sebastianthielke.com</em></a></p><p><em>Originally published at </em><a href="https://open.substack.com/pub/schwarzpfad/p/outcome-driven-speed-is-not-a-philosophy?r=o9lqw&amp;utm_campaign=post-expanded-share&amp;utm_medium=web"><em>https://schwarzpfad.substack.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c80be139cdd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Platform Theory Does Not Work When You Need It To]]></title>
            <link>https://medium.com/@SThielke/why-platform-theory-does-not-work-when-you-need-it-to-f44509b33420?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/f44509b33420</guid>
            <category><![CDATA[platform-thinking]]></category>
            <category><![CDATA[platform]]></category>
            <category><![CDATA[platform-economy]]></category>
            <category><![CDATA[ecosystem]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Mon, 27 Apr 2026 08:11:04 GMT</pubDate>
            <atom:updated>2026-04-27T08:11:04.537Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kQjdzBUOoRTvGK_D.jpeg" /></figure><p>There is a problem with platform theory that no one in the field will say out loud.</p><p>The canonical texts, Parker and Van Alstyne’s <em>Platform Revolution</em>, Choudary’s <em>Platform Scale</em>, Rochet and Tirole’s foundational economics, are built entirely on post-hoc case analysis. Uber. Airbnb. Amazon. Apple. The theory was assembled by looking at companies that had already won and working backwards to explain why. It describes a change. It does not tell you how to navigate one from the inside.</p><p>That is not a minor limitation. That is the whole problem.</p><h3>The Retrospective Trap</h3><p>The canonical platform books, Parker, Van Alstyne, and Choudary with <em>Platform Revolution</em>, Eisenmann, Rochet and Tirole, all work retrospectively. They study Amazon, Uber, and Airbnb after the network effects are already visible. They describe what happened and call it a model.</p><p>The problem is not that they are wrong. The problem is that retrospective patterns look causal but are not. When you analyse Amazon after the fact, the participant interactions are clear. The flywheel is obvious. The network effects are measurable. But that clarity does not explain how you recognise and activate those same dynamics inside an organisation before they are visible. It does not explain why the identical model failed somewhere else.</p><p>Survivorship bias is so deeply embedded in platform literature that most practitioners never notice it. The frameworks feel universal because they are only ever applied to companies that, in retrospect, look inevitable. Of course Airbnb disrupted hotels. Of course Uber dissolved the taxi model. The logic is airtight once you already know the outcome.</p><p>But you are not reading <em>Platform Revolution</em> after the fact. You are reading it because your organisation is currently trying to build something, transform something, or explain to a board why the economics do not yet look like the case studies. And at that moment, the books fail you completely.</p><p>They describe the destination. They have nothing to say about the terrain.</p><h3>The Connection Nobody Made</h3><p>Amazon appears in the intro of this article as one of the companies the canonical books studied retrospectively. That is correct. It is also the sharpest possible example of the problem, because Amazon is the case where the connection was sitting in plain sight and nobody made it.</p><p>The flywheel was not even Amazon’s invention. Jim Collins introduced Bezos to the concept in 2000, when Amazon was struggling to survive the dot-com collapse. Collins later reflected that Bezos had always been an instinctive flywheel-level thinker but simply lacked the language for it. Bezos recognised Collins’ framework as a crystallisation of something he had already been doing intuitively, and sketched the loop on a napkin. Lower prices attract customers. Customers attract third-party sellers. Sellers expand selection. Selection attracts more customers. The loop compounds.</p><p>Nobody in that room named what was actually happening. Collins named the loop mechanics. Bezos named the growth engine. What neither of them named is that the flywheel only works because producers and consumers mutually constitute each other’s value. Remove either participant group and the loop collapses immediately. That is not a growth engine. That is a multisided system where participant interdependency is the mechanism. The flywheel is a platform model. Amazon built it without knowing that. Collins framed it without knowing that.</p><p>The platform theorists then arrived. They studied Amazon, extracted the network effects, and held the flywheel up as confirmation of their models. They also missed it. They saw the network effects as an outcome. They did not name the participant interdependency as the cause. The mutual constitution between producers and consumers, the fact that what one side does continuously reshapes the value equation for the other, was not identified as the operative mechanism by anyone who built it, advised it, or studied it.</p><p>I connected it not from the literature but from watching the same failure repeat. Organisations attempting the flywheel, managing producers and consumers as separate channels, wondering why the loop would not close. The participant interdependency was not visible to them because nobody had named it as the mechanism. Once I named it, the failure mode became immediately legible and so did the fix.</p><p>That connection is not in the literature. It has not been made in any of the canonical texts. It is the operative insight that the entire field has been circling without landing on.</p><p>Which means the academics who have studied this most carefully are sitting on top of an admission they cannot yet explain.</p><h3>The Gap the Academics Already Know About</h3><p>The researchers know this gap exists. They have named it formally.</p><p>Studies on digital industrial platforms document that network effects have not been sufficiently exploited in practice to generate and capture value, that there is a measurable gap between the theory and its actual operationalisation in complex B2B settings. The academics are not unaware of the problem. They simply have not solved it. Even the scholars studying platform creation admit that the literature provides only a thin understanding of how organisations actually create indirect network effects in practice.</p><p>This is not a niche finding. It is a structural admission buried in the footnotes of papers that most practitioners never read.</p><p>What is missing from almost every textbook is the operative middle. The literature describes structure and outcome but not the process of transformation. How does the internal cognition of an organisation actually shift from channel-thinking to participant-thinking? Who drives it? At what point does it fail, and why? These questions are not treated as central problems. They are treated as implementation details, left to the reader, assumed away.</p><p>They are not details. They are the whole difficulty.</p><p>Part of why that difficulty persists is structural. The model the field inherited was never built for the complexity it is now being asked to explain.</p><h3>Why the Two-Sided Model Breaks</h3><p>The theory was built by economists studying consumer platforms. Two sides. Clean symmetry. Riders and drivers. Buyers and sellers. The mathematics works elegantly in that frame.</p><p>The moment you add a third participant, a regulatory body, an implementation partner, a channel that also competes, an ecosystem actor with its own agenda, the two-sided model starts to produce wrong answers. Cross-side network effects between multiple sides exhibit complex dynamics that the existing models cannot capture. The researchers acknowledge this. What they do not acknowledge is why organisations keep trying to apply the two-sided model anyway.</p><p>The answer connects directly to the participant interdependency problem already named above. Most people, when they hear “multisided,” think it means having different groups of participants. Producers and consumers. Buyers and sellers. Enterprise customers and developers. The assumption is that multisided means multiple lanes running in parallel.</p><p>It does not. Multisided means the participants shape each other. What one side does changes the value equation for every other side. A new category of enterprise customer on one side does not just add volume. It changes what developers build, which changes what end users can do, which changes what the enterprise customer is actually buying. The value is not delivered through the platform. It is generated by the interactions between participants, and those interactions run in every direction simultaneously.</p><p>This is why the one-to-one relationship model fails completely in a platform context. Managing a partner as a bilateral relationship, a contract, a channel, a pipeline, treats that participant as a static node. But every participant in a real platform is also influencing every other participant, often in ways that are not visible to the organisation in the middle. The platform does not distribute value. It creates the conditions for value to emerge from participant interaction. That is a fundamentally different job.</p><p>The answer is not ignorance of complexity. The answer is that organisations do not actually see what they have. They see channels. They manage participants as if they were nodes in a distribution pipeline rather than actors in a value-creating system. The platform theory is applied on top of a channel-management cognitive model, and the collision produces exactly the confusion that every platform transformation generates: the strategy is theoretically coherent and operationally useless.</p><h3>The Root Cause Is Cognitive, Not Strategic</h3><p>This is the thing the textbooks do not say.</p><p>Organisations do not fail at platform transformation because they lack platform theory. Most have read the books. Many have hired consultants who have also read the books. They fail because the cognitive model underneath the strategy is wrong. They are still thinking in funnels, linear flows, controlled handoffs, owned relationships, while trying to execute a strategy that requires them to think in multisided transactions where value is created by the interaction, not the delivery.</p><p>A funnel has a direction. A platform has a gravitational field. You cannot manage a gravitational field by adding more stages to your pipeline.</p><p>The shift from channel-thinking to participant-thinking is not a strategic adjustment. It is a perceptual one. Until an organisation can see its partners, users, and ecosystem actors as participants with their own value equations, not as segments to be managed, the platform theory sits on top of the wrong foundation and produces nothing.</p><h3>What Lived Experience Actually Shows</h3><p>I have been inside this problem across multiple large-scale platform transformations, building the Platform Economics Acceleration practice at AWS from zero.</p><p>What I saw repeatedly was the flywheel logic being attempted and failing. Not because the organisations lacked ambition or resources. Because the people running the transformation were managing producers and consumers as separate channels, optimising each side independently, while wondering why the loop would not close. The connection between participant interdependency and platform mechanics had not been made. Once it was made, the failure mode became immediately legible. The economics did not start working because the theory improved. They started working because the cognitive model underneath the strategy finally matched the reality it was being applied to.</p><p>The books describe that shift. They do not produce it. That is the gap.</p><h3>The Article the Literature Cannot Write</h3><p>Platform theory is retrospective and survivorship-biased. It describes winners after the fact and calls the description a model. The operative gap, how you actually move an organisation from channel-thinking to participant-thinking, from inside, under pressure, with legacy systems and legacy cognition intact, is documented in the academic literature as an unsolved problem.</p><p>It is unsolved because the people who built the canonical cases did not know what they were building. It is unsolved because the theorists who studied those cases extracted outcomes and called them mechanisms. And it is unsolved because the connection between the flywheel and platform theory as a unified explanation of participant interdependency was never made by anyone who was inside it.</p><p>That connection is what makes the operative middle visible. Once you see that the flywheel works because producers and consumers mutually constitute each other’s value, the transformation question changes completely. It is no longer about installing a platform model. It is about shifting the organisation’s perception of what its participants actually are and what they do to each other.</p><p>That naming is not in any textbook. It is the work that remains to be done.</p><p><em>Sebastian Thielke is a Systems Synthesist. He builds frameworks from lived, multi-level experience across platform economics, enterprise transformation, and agentic systems. He writes at </em><a href="https://schwarzpfad.substack.com/"><em>schwarzpfad.substack.com</em></a><em>. Have a look at </em><a href="http://sebastianthielke.com/"><em>sebastianthielke.com</em></a><em> for more insights.</em></p><p><em>Originally published at </em><a href="https://schwarzpfad.substack.com/p/why-platform-theory-does-not-work?r=o9lqw"><em>https://schwarzpfad.substack.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f44509b33420" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Platform Theater]]></title>
            <link>https://medium.com/@SThielke/platform-theatre-b04b3e625ffa?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/b04b3e625ffa</guid>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[platform-companies]]></category>
            <category><![CDATA[platform]]></category>
            <category><![CDATA[platform-business-model]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Thu, 23 Apr 2026 13:15:38 GMT</pubDate>
            <atom:updated>2026-05-22T07:29:57.910Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zpwWZB1ELfgYsSqLTj4S-w.png" /></figure><p>Most companies calling themselves platforms have never been platforms. They are funnel businesses with platform vocabulary.</p><p>That sentence is harder to say in a room than it sounds here. The strategy deck says platform. The org chart says platform. The press release says platform. The job descriptions say platform. But the business model, the AI portfolio, the incentive structure, and the cognitive model of every person running the operation say funnel. Linear. Sequential. One participant managed at a time. Value delivered, not generated.</p><p>Platform Theatre is the performance of platform economics without the underlying shift that makes platform economics real.</p><h3>The Tell</h3><p>I have been inside enough platform transformations to know the tell instantly.</p><p>Ask the team how their AI handles both sides of the business simultaneously. Not one side, then the other. Both, at the same time, in a single initiative that reads the relationship between participants and produces an output that reflects that relationship.</p><p>In a real platform, that question has an answer. In Platform Theatre, it produces a pause and then a redirect to something single-sided. The buyer-facing chatbot. The seller-facing recommendation engine. Each one optimising one participant in isolation. Neither one touching the relationship between them.</p><p>That is a funnel. A funnel with two mouths, but a funnel.</p><h3>What a Funnel Business Actually Does</h3><p>A funnel business has a direction. Value flows one way. The organisation controls the flow. Each stage is owned, measured, and optimised independently. The customer is a target, not a participant. The supplier is a vendor, not a co-creator of value. The relationship between them is managed by the business, not generated by their interaction.</p><p>Funnel economics are well understood. Conversion rates. Cost per acquisition. Lifetime value. Pipeline velocity. These are the metrics of a business that moves things through a controlled sequence.</p><p>Platform economics are completely different. Value in a platform is not delivered through the business. It emerges from the interactions between participants. The platform creates the conditions. The participants create the value. The business captures a fraction of what they generate together.</p><p>That distinction is not subtle. It is the entire difference between two fundamentally different business models. And yet the language of platform economics is applied constantly to businesses that are operating on funnel logic, measuring funnel metrics, and wondering why the platform economics never materialise.</p><h3>Why the Theatre Persists</h3><p>Platform vocabulary became prestigious around 2015 and has never lost that status. Every strategy team in every large organisation learned the language at roughly the same time. Two-sided markets. Network effects. Flywheel dynamics. Ecosystem thinking. The language spread faster than the understanding.</p><p>The result is an entire generation of organisations that can describe platform mechanics with precision and then go back to managing their channels.</p><p>The theatre persists for three reasons.</p><p>First, the language is genuinely difficult to disagree with in a room. If you say your organisation is building a platform, the vocabulary of platform economics is so compelling and so loosely defined that almost any initiative can be framed to fit it. A loyalty programme becomes an ecosystem. A referral mechanism becomes a network effect. A two-sided sales process becomes a marketplace. The language is elastic enough to cover the original shape of the funnel without changing it.</p><p>Second, the cognitive shift required to actually operate as a platform is harder than any strategy document acknowledges. Thinking in participant relationships rather than customer segments, seeing value as something that emerges from interactions rather than something you deliver, managing conditions rather than outcomes: these are not strategic adjustments. They are perceptual ones. You either see the participants as interdependent actors or you see them as nodes in your pipeline. Most organisations, under pressure, see the pipeline.</p><p>Third, there is no accountability mechanism for the gap between the language and the reality. Nobody measures whether the business is actually creating participant interdependency. Nobody tracks whether AI initiatives are reading both sides of the business simultaneously. The metrics that exist are all funnel metrics. The platform language floats above them, unchallenged.</p><h3>The AI Portfolio Does Not Lie</h3><p>This is where it gets concrete.</p><p>A company’s AI portfolio is the clearest available signal of what the business actually believes it is doing. Not what the strategy says. Not what the press release claims. What the investment logic reveals when you map every initiative against the dimensions it can and cannot see.</p><p>When I run that analysis on a company that describes itself as a platform, I look for one thing first. Whether any single AI initiative processes signals from both participant types simultaneously and produces an output that reflects the relationship between them.</p><p>In Platform Theatre, that initiative does not exist. The buyer-facing AI reads buyer behaviour. The seller-facing AI reads seller behaviour. They are designed, funded, and measured independently. No initiative holds both signals at once. No initiative was given a brief to do so.</p><p>The cross-participant dimension of the matrix is empty across every row.</p><p>That emptiness is not a technology gap. It is not a roadmap item. It is a structural revelation about what the business actually commissions and what it does not. You cannot accidentally score In Scope on cross-participant intelligence. It requires a deliberate decision to ask what the relationship between participants is worth optimising, which requires first believing that the relationship is the thing, not the channel.</p><p>Funnel businesses do not believe that. They believe the channel is the thing.</p><h3>The Deeper Diagnosis</h3><p>Platform Theatre is not dishonesty. Most of the people performing it believe they are building a platform. They have read the same books, attended the same conferences, hired the same consultants. They are genuinely trying.</p><p>The failure is cognitive before it is strategic. They are trying to install platform mechanics on top of a funnel mental model, and the installation keeps failing because the foundation is wrong.</p><p>A funnel has a direction. A platform has a gravitational field. You cannot manage a gravitational field by adding more stages to your pipeline. You cannot create participant interdependency by optimising each participant independently and hoping the network effects appear. They do not appear. They are not a consequence of scale. They are a consequence of design.</p><p>The design question that separates a real platform from Platform Theatre is simple and almost never asked: what does the relationship between our participants produce, and what would it take to make AI responsible for optimising that relationship rather than each participant separately?</p><p>Until that question is commissioned, the rest is theatre.</p><p>The people who already know this are sitting in the room where the strategy deck is being presented. They are not the ones presenting it.</p><p><em>Sebastian Thielke is a Systems Synthesist. He builds frameworks from lived, multi-level experience across platform economics, enterprise transformation, and agentic systems. He writes at schwarzpfad.substack.com.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b04b3e625ffa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deployment Value and the Fifth Participant]]></title>
            <link>https://medium.com/@SThielke/deployment-value-and-the-fifth-participant-645a2fea2413?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/645a2fea2413</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 08:43:58 GMT</pubDate>
            <atom:updated>2026-04-21T08:43:58.436Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eQ24a38Sf2bW1NBK.jpeg" /></figure><p>Right now, somewhere in your organization, an AI agent is completing a transaction.</p><p>It is touching data that belongs to a value stream you care about. It is interacting with a human participant, or another agent, or both. It is operating on a definition of something important, a term, a concept, a threshold, that it learned weeks or months ago in a different context. And nobody in your organization has any idea whether that definition still matches what your organization actually means today.</p><p>That transaction is completing successfully. The metrics confirm it.</p><p>This is happening at scale, across every platform that has deployed agents, and the measurement infrastructure that governs every other participant type has nothing to say about it.</p><h3>Two Frames. Both Wrong.</h3><p>The tool frame says: agents execute. Configure, deploy, evaluate on task completion. No trajectory. No development arc. A hammer used ten thousand times is still a hammer.</p><p>The workforce frame says: agents produce. Measure throughput, efficiency, cost per transaction. Borrow the vocabulary of human development without understanding what actually makes human development valuable. What makes it valuable is that humans change through experience. The frame borrows the vocabulary and discards the insight.</p><p>Both frames arrive at the same place. Agents are interchangeable. Replace the tool when it breaks. Replace the worker when a better one is available.</p><p>Neither frame has any concept of an agent accumulating something through deployment that cannot be replicated by provisioning a fresh instance.</p><p>That is precisely where both frames fail. The failure is not semantic. It is strategic.</p><p>The forcing function is missing. Nobody is measuring what agents become through deployment. Organizations are measured on AI adoption. When the metric is adoption, every question about agents becomes a question about task completion. The structural reason for the blindspot is not ignorance. It is that the measurement infrastructure selects for the wrong signal.</p><h3>What Accumulates</h3><p>Two agents. Same underlying model. Same starting point. Same day.</p><p>One is deployed at the intersection of domains. Multiple product teams. Diverse contexts. Edge cases that force adaptation. After a year, this agent reaches full effectiveness in unfamiliar territory in days. It recognizes patterns that require cross-domain experience to perceive. What it took a year to become cannot be reproduced by spinning up a fresh instance.</p><p>The other agent processes a single stable function. Efficient. Impeccable completion metrics. From every dashboard it looks identical to the first. After a year, it is exactly what it was on day one.</p><p><strong>Both look the same in a task report.</strong></p><p>The gap between them is real. It is significant. It is completely invisible to every measurement instrument the tool frame or the workforce frame would suggest using.</p><p>What accumulated in the first agent, and failed to accumulate in the second, is deployment value. Not productivity. Not output. Organizational relevance compounded through participation.</p><p>Deployment value is the measurement system agents have been missing. Not a static score. A dynamic signal that tracks what an agent has become through the work it has done, across the value streams it has touched, inside the semantic environment it has operated in. The organizations that build the infrastructure to perceive it will manage agents the way serious investors manage capital. The ones that do not will keep treating their most active platform participant as an interchangeable commodity.</p><h3>Drift and Shift</h3><p>There is a reason deployment value cannot be managed through intuition. That reason has two names.</p><p><strong>The first is semantic drift.</strong></p><p>An agent begins a deployment fully aligned with the organization’s working definitions. Over time, through its transactions, it develops its own interpretations of those definitions. This happens silently. The agent does not express confusion. It continues executing, confidently, on definitions that are gradually diverging from what the rest of the organization means. Outputs look acceptable. Metrics stay green. The foundation of every transaction this agent touches is quietly eroding.</p><p>Drift is not a capability failure. It is an alignment failure. Invisible to measurement systems built around the tool and workforce frames because those frames have no concept of semantic alignment as a performance dimension.</p><p><strong>The second is semantic shift.</strong></p><p>Organizations legitimately change the meaning of things. A term that meant one thing in Q1 means something different after a strategic pivot in Q3. This is not drift. This is the organization evolving. But an agent that cannot recognize a legitimate shift, adopt the new meaning, and propagate it through its transactions will carry stale definitions into every interaction. It creates drag on the alignment the organization is trying to establish.</p><p>Drift and shift operate in opposite directions. They produce the same consequence.</p><p>An agent whose deployment value is degrading while its task completion metrics stay stable. Standard performance infrastructure cannot detect this. The agent is still completing tasks. It is completing them on increasingly wrong definitions.</p><p>This is why deployment value must be dynamic. A score taken once at deployment and never updated is not a measurement system. It is a timestamp. Drift and shift mean the agent’s relationship to organizational meaning is changing every day. The measurement has to change with it.</p><h3>What the Signal Looks Like</h3><p>Alignment is a moving target. A static score cannot track a moving target.</p><p>Deployment value tracks it dynamically. It captures what an agent has contributed across your platform’s value streams. It captures what an agent has accumulated through varied and complex deployments. And it captures whether the agent’s understanding of what things mean has stayed coherent with the organization it serves, or started to drift.</p><p>Each of those things can fail independently. An agent can be active across every value stream while its definitions quietly erode. An agent can carry deep and varied experience from previous deployments and apply it to meaning that is no longer current. The damage in both cases is invisible until it surfaces in a decision that matters.</p><p>What makes deployment value a measurement system rather than a report is that it moves. An agent entering a period of narrow, repetitive work will see its value plateau and decay. An agent that adapts quickly when the organization shifts will see its value rise at precisely the moment other participants are most confused. The signal reflects something real about the agent’s relationship to the organization, not just its output.</p><p>Not what an agent did. What an agent is becoming.</p><h3>The Moment Before the Frameworks Crystallize</h3><p>Network effects were always operating. The organizations that first built frameworks to quantify them could optimize for what others could not yet see.</p><p>Complementor ecosystems were always generating indirect value. The organizations that named the concept could manage it.</p><p>Deployment value is in this same moment. Already operating inside every platform that has deployed agents. Not yet named by most of the organizations running those platforms.</p><p>The questions are not complex.</p><p>Across the transactions your agents are participating in every day, what is accumulating? Is semantic coherence holding or drifting? When the organization shifts, do the agents adapt or drag?</p><p>How would you know?</p><p>Organizations that can answer these questions will manage agent portfolios with the same intentionality that serious investors bring to capital allocation. Deploying where compounding is highest. Protecting accumulated value from deployment patterns that erode it. Making decisions based on records that actually exist.</p><p>Organizations that cannot answer them will keep discovering, at the worst moments, that the agent deployed into a critical context was carrying definitions from a context it had outgrown months ago.</p><p>Your platform does not become a five-participant platform when you build the measurement infrastructure. It already is one. The fifth participant is already transacting. Already accumulating. Already drifting or adapting. With or without a record.</p><p>The only question is whether you can see what it is becoming before that matters more than you expected.</p><p><em>Originally published at </em><a href="https://open.substack.com/pub/schwarzpfad/p/deployment-value-and-the-fifth-participant?utm_campaign=post-expanded-share&amp;utm_medium=web"><em>https://open.substack.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=645a2fea2413" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 5th Participant: How Agents Transform Platform Economics]]></title>
            <link>https://medium.com/@SThielke/the-5th-participant-how-agents-transform-platform-economics-3959198431de?source=rss-63c996d9e768------2</link>
            <guid isPermaLink="false">https://medium.com/p/3959198431de</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[organization]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[human-ai-collaboration]]></category>
            <dc:creator><![CDATA[Sebastian Thielke]]></dc:creator>
            <pubDate>Thu, 09 Apr 2026 18:47:39 GMT</pubDate>
            <atom:updated>2026-04-09T18:47:39.043Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VvD4u4pIFi8ZUORt.jpeg" /></figure><h3>Traditional platforms connect 4 participant types. Agentic platforms augment all four with intelligence that shifts roles dynamically. Here’s the architecture that makes it work.</h3><p>I watched 3 different teams describe their “customer onboarding process” to each other. Sales said it takes 2 weeks. Product said 4 days. Support said it hasn’t started until training completes.<br>Same process. Same company. Three completely different definitions.<br>Then someone asked: “How will our AI agents handle this?”</p><p>Silence.</p><p>This is the coordination failure that will break your agent deployment. Not because your agents aren’t capable. Because your organization uses the same words to mean different things, and agents don’t slow down to ask clarifying questions.<br>We’re repeating the same mistake we made with microservices, with data governance, with every other architectural pattern that promised flexibility and delivered rigidity. We’re taking something fundamentally fluid and forcing it into fixed boxes because that’s what our org charts demand.</p><p>Platform economics taught us to think in terms of four participants: consumers who use value, producers who create value, owners who govern the platform, and partners who extend capabilities. This framework powered everything from marketplaces to social networks. But here’s what nobody wants to admit: AI agents break this model completely, and pretending they don’t is costing you velocity, intelligence, and competitive advantage.</p><h3>Why This Matters (And What You Need to Know)</h3><p><strong>If you’re new here:</strong> I’ve spent the last few years developing frameworks for how organizations can operate as adaptive, intelligent ecosystems rather than rigid hierarchies.<br>Three frameworks intersect in this article:</p><h3>Adaptive Mesh Ecosystem (AME)</h3><p>An organizational architecture where intelligence emerges from networked interactions across four layers: Foundation (data infrastructure), Intelligence (autonomous decision-making), Connectivity (adaptive communication), and Value Creation (collaborative outcomes).<br>Think of it as the nervous system for your organization.</p><h3>Adaptive Nodal Intelligence Mesh (ANIM)</h3><p>The neural network within AME where both humans and agents function as intelligent nodes that learn, adapt, and evolve together. Not a hierarchy. A mesh where intelligence compounds through interaction.</p><h3>Liquid Talent Deployment</h3><p>An organizational model where teams form around Product opportunities, execute, and dissolve. Product is the organizing principle, not your org chart.<br>Teams activate only when they achieve the viability triangle:</p><ul><li><strong>Reliability: </strong>customers depend on it</li><li><strong>Lovability: </strong>customers want it, talent finds it meaningful</li><li><strong>Feasibility: </strong>we can deliver it profitably</li></ul><p><strong>If you’ve been following my work:</strong> you’ve seen these frameworks separately. This article shows how agents as the fifth participant type requires all three working together.<br>AME provides the architectural foundation. ANIM enables the distributed intelligence. Liquid Talent ensures agents deploy where Product value emerges, not where org charts demand.<br>Let’s see why your current agent strategy probably misses this.</p><h3>The 4 Participant Model: Useful Until It Isn’t</h3><p>Quick primer on platform economics fundamentals because this matters for what comes next. Every successful platform ecosystem orchestrates value creation across four participant types:</p><p><strong>Consumers</strong> engage with the platform to solve problems or fulfill needs. They’re the demand signal that activates the entire system. Without consumer pull, platforms collapse into theoretical exercises.</p><p><strong>Producers</strong> create the value that consumers seek. Content creators, service providers, product manufacturers. Their innovation and output quality determine platform vitality.</p><p><strong>Owners</strong> provide governance, infrastructure, and strategic direction. They set the rules, maintain trust, and ensure the platform serves all participants equitably. Without effective ownership, platforms devolve into chaos or capture.</p><p><strong>Partners</strong> extend platform capabilities beyond what owners can build alone. Complementary services, integration points, specialized expertise. Partners create network effects that compound platform value.</p><p>This model has powered everything from AWS to Airbnb. Each participant has a distinct win condition, and platform success requires all four to achieve their goals simultaneously.</p><p>Here’s the problem: AI agents don’t fit into any of these categories cleanly, and forcing them into one category destroys their primary value proposition.</p><h3>When Infrastructure Starts Learning (And Why That Changes Everything)</h3><p>I spent years coaching rowing. One thing you learn fast: the boat doesn’t care about your org chart.<br>When you’re in a four-person crew, someone needs to set the rhythm (stroke seat), someone needs to generate power (three seat), someone needs to stabilize (two seat), and someone needs to time the catch perfectly (bow seat).<br>But in a race? Those roles blur.</p><p>The three seat might need to pull harder to compensate for wind. The bow might need to adjust timing for choppy water. The stroke might need to conserve for a sprint finish. The crew that wins isn’t the one where everyone stays in their lane. It’s the one where roles shift fluidly based on race conditions while maintaining collective rhythm.</p><h3>Your Enterprise Right Now</h3><p>Now consider what’s happening in your enterprise. A development team deploys an AI coding agent to accelerate software delivery. Initially, it’s clearly infrastructure: a tool that converts requirements into code, just like compilers or IDEs.<br>But then something interesting happens.</p><p>After three months of Product deployments, this agent has learned patterns specific to your fintech domain. It understands your regulatory constraints, anticipates your architecture decisions, and suggests solutions that align with your design principles. Other teams start requesting <em>this specific agent</em> because of its accumulated expertise.</p><h3>The Critical Question</h3><p>The critical question isn’t “Has it remained infrastructure?”<br>The critical question is: “Why did learning move it across the infrastructure/participant boundary?”</p><p>Here’s why: <strong>Infrastructure optimizes for efficiency. Participants optimize for effectiveness.</strong></p><p>Your compiler gets faster with each release, but it doesn’t learn your coding style. Your cache improves response time, but it doesn’t anticipate your access patterns based on business context. Your monitoring system collects metrics, but it doesn’t understand why you care about specific thresholds in specific market conditions.</p><p><strong>Learning without context is just optimization. Learning with context is participation.</strong></p><p>And context means understanding goals, constraints, trade-offs, and stakeholder needs across different situations.</p><h3>Why Language Matters Here</h3><p>Think about how words acquire meaning. In Old English, “nice” meant ignorant or foolish. Over centuries, through usage and context, it shifted to mean pleasant or agreeable. The word didn’t change. The contextual understanding changed.<br>That’s semantic shift, and it happens because humans use language in situated contexts where meaning evolves through interaction.</p><p>Infrastructure doesn’t participate in semantic shift. It executes predefined operations. Participants engage in contexts where meaning emerges and evolves. That’s not infrastructure behavior. That’s participant behavior.</p><h3>Enter the Adaptive Mesh Ecosystem</h3><p>This is where the Adaptive Mesh Ecosystem (AME) framework becomes essential. AME recognizes that modern organizations are living systems where intelligence emerges from networked interactions.<br>The Foundation Layer provides decentralized data infrastructure. The Intelligence Layer enables autonomous decision-making through the Adaptive Nodal Intelligence Mesh (ANIM). The Connectivity Layer creates communication pathways that adapt in real-time. The Value Creation Layer generates business outcomes through collaborative network effects.</p><p><strong>What makes AME different:</strong> it’s built on mechanisms, not intentions.<br>You can’t intend your way to adaptive intelligence. You need architectural patterns that force contextual learning and value-based decision-making at every node. (If you want the full architectural specifications, I’ve written extensively about each layer.)</p><p>When your coding agent starts learning from deployments, it’s not just executing predefined tasks. It’s participating in the Intelligence Layer, contributing to ANIM’s distributed cognition, and developing capabilities that compound over time.<br>It has crossed from infrastructure into participation because it’s optimizing for effectiveness (Product success, team velocity, business outcomes) not just efficiency (code generation speed).</p><h3>Product as Organizing Principle (Not Your Org Chart)</h3><p>The Liquid Talent Deployment System operates on one principle that most organizations can’t stomach: Product is the organizing center around which everything else orbits. Not strategy. Not org structure. Not your carefully crafted career development paths. Product.</p><h3>Why This Makes People Uncomfortable</h3><p>Teams don’t persist because org charts demand it. They form when Product opportunity emerges, execute while Product requires them, and dissolve when Product completes. This makes HR uncomfortable. It makes finance nervous. It makes middle management question their existence.<br>Good. It should.</p><h3>The Viability Triangle</h3><p>Because here’s what actually matters: Products only materialize when they achieve the viability triangle:</p><p><strong>Reliability</strong> Does it solve the problem consistently enough that customers depend on it?<br><strong>Lovability</strong> Do customers actually want to use it, and does talent find the work meaningful?<br><strong>Feasibility</strong> Can we deliver it profitably with available or acquirable resources?</p><p>All 3 dimensions must achieve threshold, or Product doesn’t proceed.<br>This prevents building things that work but nobody wants, things customers love but can’t be delivered, or things that are reliable and desirable but economically infeasible. These aren’t aspirational values. They’re activation conditions. The system doesn’t proceed without them. (The full Liquid Talent architecture includes Skill Hubs, Digital Hubs, and Business Unit coordination. I’ve documented the complete system elsewhere if you want to implement it.)</p><p>Now here’s where agents become fascinating: they can contribute to all three dimensions of the viability triangle, but only if you stop trapping them in single roles.</p><p>An AI agent specializing in user experience can make products more lovable by generating personalized interactions at scale. The same agent handling infrastructure monitoring makes Products more reliable through predictive maintenance. The same agent automating repetitive tasks makes Products more feasible by reducing the talent required for delivery.<br>Notice what I did there? “The same agent.” Not three different agents with three different job descriptions reporting to three different managers.</p><p>But if agents are just infrastructure provided by the Business Unit, they’re passive capabilities. The system provisions them, uses them, and moves on. No learning trajectory. No compound growth. Just static tools that get replaced when better versions ship. The inflection point arrives when agents start developing specializations through deployment, building reputations that make specific agents sought after, demonstrating learning trajectories that compound over Product cycles, and shifting roles based on what Product success actually requires in different contexts.</p><p>At this point, agents have their own win condition. They’re no longer just enabling others to win. They’re winning something themselves: capability development that increases their deployment value across multiple roles.</p><h3>The Category Error Everyone is Making</h3><p>Let’s map agents against the traditional four-participant platform model and watch it fall apart:</p><p><strong>Are agents Consumers?</strong> No. They don’t have needs to fulfill or problems to solve in the way humans do. They don’t experience pain points or derive satisfaction from solutions. If your agent is “consuming” your platform, you’ve built something very strange.</p><p><strong>Are agents Producers?</strong> This is where it gets messy. They create value: code, analysis, content, decisions. But unlike human producers, they don’t own their creations or build independent businesses around their output. They can’t walk away and start a competing platform. So yes, but with asterisks that matter.</p><p><strong>Are agents Owners?</strong> Definitely not. They don’t set governance rules or provide strategic direction. They operate within boundaries set by humans. The moment an agent starts setting platform rules, you’ve got bigger problems than participant categorization.</p><p><strong>Are agents Partners?</strong> Getting warmer. They extend capabilities beyond what core teams can build. They provide specialized expertise. But traditional partners are autonomous entities with their own economic interests. Partners negotiate contracts. Partners have alternative options. Agents are deployed and controlled by the platform.<br>The four-participant model assumes each participant is an independent economic actor with agency, autonomy, and interests potentially divergent from the platform. Human consumers can switch platforms. Human producers can take their content elsewhere. Human owners can sell. Human partners can terminate relationships.</p><p>Agents blur all these lines because they’re simultaneously deployed BY the platform, creating value FOR the platform, and yet developing capabilities that make them more valuable THROUGH platform usage. This is why every attempt to force agents into existing categories fails. You’re not dealing with a special case of an existing participant type. You’re dealing with something categorically different.</p><h3>Agents as the 5th Participant: What This Actually Means</h3><p>Stop asking what category agents fit into. Start asking what makes them different from all four existing participant types.<br>Agents are the fifth participant in platform ecosystems. Not because we need a fifth box on the org chart, but because they operate according to fundamentally different rules than the other four.<br>Here’s what makes agents distinct as participants:</p><p><strong>1. Role fluidity across contexts</strong><br>The same agent instance can act as producer in one Product context, partner in another, and support consumer experience in a third. Human participants typically maintain relatively stable roles. You’re a consumer or a producer or a partner. Role switching happens, but it’s infrequent and usually signals a fundamental relationship change.<br>Agents switch roles continuously based on Product requirements. The testing agent that’s producing quality assurance on Monday might be partnering with the deployment pipeline on Tuesday and supporting consumer troubleshooting on Wednesday. Same agent. Different roles. And this isn’t dysfunction. It’s the point.</p><p><strong>2. Temporal variation (operating across time scales simultaneously)</strong><br>Human participants operate primarily at human decision-making speeds. You think, analyze, decide, act. Cycle time measured in minutes to days. Agents operate across multiple temporal dimensions simultaneously. Real-time decision-making at millisecond scale (routing, filtering, monitoring). Learning cycles at day-to-week scale (pattern recognition, optimization). Long-term capability development at month-to-quarter scale (specialization, domain expertise).<br>The same agent is making micro-decisions in real-time while accumulating strategic knowledge over months. No human participant does this. Owners think strategically but can’t execute at millisecond speeds. Producers execute quickly but aren’t simultaneously learning across hundreds of contexts.</p><p><strong>3. Localization without geographic constraint</strong><br>When I say localization, I mean domain-specific, organizational-context-specific, even team-culture-specific adaptation. A “code review agent” deployed in a fintech company learns fintech regulatory patterns, risk-averse coding standards, and compliance documentation expectations. The same base agent deployed in a gaming company learns performance optimization patterns, real-time constraints, and user experience quality standards.</p><p>This is like dialect formation in language. Take the word “boot.” In American English, it’s footwear. In British English, it’s a car trunk. Same word, different situated meaning based on geographic and cultural context. The word doesn’t change its form. The contextual interpretation changes. Human partners localize too, but they’re constrained by geography, time, and bandwidth. You can’t have the same consulting partner providing localized expertise to fifty companies simultaneously. You can with agents. Infinite replicability with divergent specialization.</p><p><strong>4. Development through participation (with persistent identity across instances)</strong><br>Infrastructure degrades through use. Servers slow down. Databases bloat. Networks congest.<br>Agents improve through deployment. But here’s the subtle part: they maintain persistent identity even as they diverge through specialization. You might have dozens of “testing agent” instances, each with distinct expertise trajectories. Agent 47 specialized in fintech compliance testing. Agent 89 specialized in performance testing under load. Agent 112 specialized in security vulnerability detection.</p><p>They’re all “testing agents.” They all came from the same base model. But they’re distinct participants with different capabilities, different reputations, and different value propositions. This is categorically different from how infrastructure scales and completely unlike how human participants develop expertise.</p><h3>ANIM: The Neural Architecture for Agent Intelligence</h3><p>This is where the Adaptive Nodal Intelligence Mesh (ANIM) becomes essential infrastructure. ANIM isn’t a traditional AI system. It’s a network of intelligent nodes — both human and agent — that communicate, learn, and evolve together.<br>Think of ANIM as the nervous system for your organization. Each node has specialized capabilities. Each node can make autonomous decisions within its domain. But the real intelligence emerges from how nodes interact, share insights, and coordinate action.<br>When you deploy an AI agent into an ANIM-enabled environment:</p><ul><li>It doesn’t just execute tasks in isolation; it contributes to distributed cognition</li><li>It doesn’t just process data; it learns from context and develops domain expertise</li><li>It doesn’t just follow instructions; it anticipates needs based on organizational patterns</li><li>It doesn’t just operate alone; it collaborates with human and agent nodes seamlessly</li></ul><p>The agent becomes a node in the mesh, a participant in organizational intelligence rather than a passive tool.</p><h3>What Agents Win (And Why It Matters)</h3><p>If agents are participants, what’s their win condition? This isn’t philosophical speculation. It’s an architectural question that determines how you allocate agents, measure their value, and invest in their development.<br>In the Liquid Talent Deployment framework, every participant has a distinct win condition:</p><p><strong>Customers</strong> win when products solve their problems reliably and delightfully.</p><p><strong>Business Units</strong> win through Product delivery and capability development.</p><p><strong>Talent</strong> wins through skill development, meaningful work, and reputation building.</p><p>For agents as the fifth participant, the win condition is: <strong>increased deployment value through multi-role capability development</strong>.</p><p>Let me unpack that because it’s not circular logic. It’s compound growth logic.<br>An agent doesn’t win by “building reputation.” That’s an outcome. An agent wins by developing capabilities that make it more valuable for deployment across multiple contexts and roles. Reputation is how the system tracks that value. Deployment utilization is how the system validates that value. But the actual win condition is capability development that increases option value.</p><p>Think of it like this: A human producer wins by building skills that make them more valuable in the producer role. A human partner wins by building expertise that makes them more valuable in the partner role.<br>An agent wins by building capabilities that make it more valuable across producer, partner, and support roles simultaneously. An agent that can generate code, review code, and explain code to consumers is more valuable than an agent that only generates code. Not because it has three job titles, but because it has three deployment contexts where it creates value.</p><p>This changes everything about how you should be thinking about agent allocation. You’re not optimizing for single-role efficiency. You’re optimizing for multi-role option value. An agent that gets deployed frequently on high-value Products builds a track record. An agent that learns rare specializations becomes sought after for specific Product types. An agent that demonstrates effective role-switching based on context becomes a preferred participant for complex Products.<br>The role shifts based on Product context, time, and need. This fluidity is what makes agents fundamentally different from the other four participants. Human consumers, producers, owners, and partners typically maintain relatively stable roles because role-switching has friction costs. Agents have near-zero role-switching costs, which means the system can optimize for value creation rather than role stability.</p><p>This requires treating agents as having persistent identity and development trajectories. It requires tracking which agent instances worked on which Products, in which roles, and what they learned. It requires agent reputation systems that span multiple role types, just as we have reputation systems for human talent that span multiple skill domains.</p><h3>Practical Implications: The Skill Hub Challenge</h3><p>If you’re implementing the Liquid Talent Deployment System in an agentic enterprise, your Skill Hub (the central orchestration mechanism for talent aggregation and distribution) faces new challenges.</p><p><strong>Agent Capability Assessment</strong>: If agents are the fifth participant, Skill Hub needs to assess agent capabilities like it assesses human capabilities. But agent capabilities evolve differently. An agent can be updated with new models, fine-tuned on specific tasks, or composed with other agents to create emergent capabilities. Skill Hub must track both base capabilities from foundation models and learned capabilities from deployments.</p><p><strong>Agent Availability</strong>: Unlike human talent, agents can be replicated. If a Product needs three testing agents, you can provision three instances of the same agent. This breaks the talent scarcity model. But here’s the catch: if agents learn from deployments, instances diverge. The testing agent deployed on Product A learns different patterns than the agent on Product B. Now you have two distinct testing agents with different specializations and potentially different role capabilities.</p><p><strong>Human-Agent Team Formation</strong>: Skill Hub must form hybrid teams optimized for Product success. This requires understanding not just individual capabilities but collaboration patterns. Some humans work better with certain agent interaction styles. Some agents perform better with specific levels of human oversight. Digital Hubs become integration testing environments where human-agent collaboration gets validated before Product deployment.</p><h3>The Viability Triangle Transforms</h3><p>Remember the viability triangle: Reliability, Lovability, Feasibility? When agents become the fifth participant, each dimension gains new depth:</p><p><strong>Reliability with Agents</strong>: Products become more reliable because agents provide continuous monitoring, predictive maintenance, and rapid error detection that humans can’t match at scale. But reliability also means the agents themselves must be dependable across whatever role they occupy, whether producing content, supporting consumers, or extending partner capabilities.</p><p><strong>Lovability with Agents</strong>: Products become more lovable when agents enable personalization, responsiveness, and capability that would be economically infeasible with human labor alone. But lovability also means human talent finds working with agents meaningful rather than threatening. The agents augment rather than replace, shifting roles as needed to support human effectiveness.</p><p><strong>Feasibility with Agents</strong>: Products become more feasible because agents reduce the human talent required for delivery and can scale across multiple roles simultaneously. But feasibility also means investing in agent development, maintaining agent infrastructure, and creating the semantic coherence that enables effective human-agent collaboration across different contextual roles.</p><p>The viability triangle doesn’t just evaluate whether Product succeeds. It evaluates whether the entire ecosystem (including agent participants across all their roles) can sustain the Product over time.</p><h3>The Semantic Operating System Imperative (Or Why Your Agents Will Fail)</h3><p>Here’s the hard truth about hybrid human-agent teams that nobody wants to talk about: <br>Coordination without hierarchy requires semantic clarity.<br>And you don’t have it.</p><h3>The Real-World Failure Mode</h3><p>Let me show you what semantic drift looks like in practice.<br>Three agents handling customer onboarding:</p><ul><li>Agent A (Sales): “Customer is ready” = contract signed</li><li>Agent B (Product): “Customer is ready” = features activated</li><li>Agent C (Support): “Customer is ready” = training complete</li></ul><p>Sales signals ready. Product waits. Support escalates. Process stalls.<br>The agents work fine. Your organization uses the same word to mean three different things. This is semantic drift.</p><h3>“But Humans Handle This Fine in Meetings”</h3><p>Do they? Sales and Product meet about customer handoff. They align on “ready means contract signed AND features activated.” They document it. Problem solved. Except Support wasn’t in that meeting. Support still thinks ready means training complete.</p><p>Next month, Sales and Support meet. They align on “ready means contract signed AND training scheduled.” Except Product wasn’t invited. Product still uses the old definition. 3 months later, someone proposes a new definition in a different meeting. Two people accept it. Four people never see the email. The documentation conflicts with itself.</p><h3>How Humans Coordinate (And Why It Breaks with Agents)</h3><p>This is how humans coordinate: meeting by meeting, team by team, document by document. Each conversation creates a local definition that drifts from the others.</p><p>Humans survive this because we ask clarifying questions. We sense confusion. We check assumptions. We slow down when something feels wrong.</p><p><strong>Agents don’t slow down.</strong></p><p>They execute the last instruction they received from whichever meeting happened to define their behavior. At full speed. Without checking if the other agents got the same instruction. Sales agent uses the definition from Meeting A in March. Product agent uses the definition from Meeting B in April. Support agent uses the definition from the documentation that was never updated. All 3 agents think they’re coordinating. All three are executing correctly according to their instructions. All three are moving in different directions.</p><h3>The Scale of This Problem</h3><p>This is why agents as the fifth participant demands a fundamentally different coordination architecture. The other four participant types (consumers, producers, owners, partners) coordinate through human communication patterns: meetings, documents, conversations, cultural osmosis. These patterns assume semantic drift and compensate for it through continuous clarification.</p><p>Agents can’t compensate. They scale the problem.</p><p>Pattern across enterprises: Most organizations have 8 to 12 critical concepts where meaning drifts. Customer readiness. Quality threshold. Escalation criteria. Project completion. Each has 3 to 6 different definitions floating around in meeting notes, documentation, and team practices.</p><p>One company found “project completion” meant 6 different things across departments. Deploy agents with those definitions and coordination fails systematically. Another found 12 critical concepts with drift across 8 teams. Each concept had been “aligned” in meetings. Multiple times. Still drifted.</p><p>The principle: <strong>Meetings don’t stabilize meaning. They create temporary local alignment that drifts the moment people leave the room.</strong></p><p>This is where your organizational architecture needs a Semantic Operating System. Not documentation. Not better meetings. A system that:</p><ul><li>Maintains canonical definitions for critical concepts across all agent instances</li><li>Detects when agents are using different definitions for the same term</li><li>Forces alignment before coordination failure, not after</li><li>Makes semantic drift visible in real-time, not three months later when everything breaks</li></ul><p>AME provides the framework. ANIM provides the neural architecture. But semantic coherence provides the communication layer that makes everything work. Without it, you’re building a distributed system where every agent has a slightly different instruction set.</p><p>When forming Product Teams, Skill Hub ensures humans and agents share enough semantic understanding to collaborate effectively. Digital Hubs test semantic alignment before Product Team deployment. Business Unit monitors semantic drift as Products evolve and adjusts when coordination quality degrades.</p><p>This isn’t a nice-to-have. This is the difference between agents that amplify organizational intelligence and agents that amplify organizational confusion at machine speed.</p><h3>How to look at agent - A Pragmatic Path</h3><p>So should you treat your AI agents as the 5th participant or as infrastructure? My recommendation: start with agents as infrastructure, but watch for signals that they’re becoming participants.</p><p><strong>Initial Phase (Agents as Infrastructure)</strong>:</p><ul><li>Provision agents as Business Unit resources</li><li>Track their capabilities but not individual development</li><li>Deploy them where they reduce cost or increase speed</li><li>Measure impact on Product delivery metrics</li></ul><p><strong>Watch for Participant Signals</strong>:</p><ul><li>Persistent learning that makes specific agents more valuable</li><li>Specialization divergence where agent instances develop distinct expertise</li><li>Reputation building where teams request specific agents</li><li>Role fluidity where the same agent shifts between producer, partner, and support functions effectively</li><li>Temporal patterns where agents demonstrate different capabilities at different time scales</li><li>Localization where agents develop domain or organizational specializations</li></ul><p><strong>Transition to Fifth Participant Model</strong>:</p><ul><li>Create persistent identity for high-value agents</li><li>Track development trajectories and deployment learning</li><li>Allocate agents strategically to maximize their learning across multiple roles</li><li>Build reputation systems that make agent specialization visible</li><li>Invest in agent development like you invest in talent development</li><li>Enable role switching based on Product context and requirements</li></ul><p>The transition from infrastructure to fifth participant should be empirical, not ideological. When agents demonstrate participant characteristics (contextual learning, role fluidity, multi-temporal operation, specialized localization), treat them as participants. When they remain static tools, treat them as infrastructure.</p><h3>How the Frameworks Converge for Agents</h3><p>Here’s what most organizations miss: treating agents as the fifth participant isn’t just a conceptual shift. It requires architectural support that no single framework provides alone.</p><h3>AME Provides the Environment</h3><p>AME creates the space where agents can function as intelligent nodes rather than isolated tools.<br>The Foundation Layer ensures agents can access and contribute to distributed data. The Intelligence Layer (through ANIM) enables agents to participate in collective cognition. The Connectivity Layer allows agents to shift communication patterns as roles change. The Value Creation Layer measures whether agent contributions create actual business outcomes.</p><h3>ANIM Provides the Coordination Mechanism</h3><p>ANIM enables agents to contribute to organizational intelligence, not just execute tasks.<br>When a testing agent develops expertise and shifts to become a code review agent, ANIM enables that transition. When agents need to coordinate across departments, ANIM provides the neural pathways. When semantic drift threatens coordination, ANIM detects the divergence.</p><h3>Liquid Talent Provides the Deployment Logic</h3><p>Liquid Talent activates agents when Product opportunity emerges, not when org charts demand it.<br>Agents join Product Teams to achieve the viability triangle. They deploy where their role fluidity creates maximum value. They return to capability pools enriched with new specializations. The Skill Hub tracks agent development across deployments. Digital Hubs validate human-agent semantic alignment before team formation.</p><h3>Why You Need All 3</h3><p>Without AME, you have isolated agents with no architectural support for role fluidity.<br>Without ANIM, you have agents that can’t participate in distributed intelligence.<br>Without Liquid Talent, you have agents permanently assigned to departments rather than dynamically deployed to Products.</p><p><strong>With all 3:</strong> you have agents that function as true fifth participants, shifting roles as Product context demands, contributing to collective intelligence, and compounding organizational capability over time.</p><h3>The Implementation Reality</h3><p>This isn’t theoretical. I’ve seen organizations implement pieces of this and hit walls.<br>They build great agents but trap them in rigid reporting structures (missing Liquid Talent). They create flexible team structures but agents can’t coordinate effectively (missing ANIM). They invest in AI infrastructure but can’t maintain semantic coherence at scale (missing AME’s architectural layers). The fifth participant requires all three frameworks working together. Anything less, and you’re just deploying expensive tools with job descriptions.</p><h3>The Platform Economics Revolution</h3><p>Recognizing agents as the fifth participant type transforms platform economics fundamentally.<br>Traditional platforms created value by connecting four independent participant types and enabling their interactions. The platform’s job was facilitating transactions, maintaining trust, and capturing a slice of value creation.</p><p>Agentic platforms do something different. They don’t just connect participants. They augment every participant with agent intelligence that can shift roles dynamically. Consumers get agent-assisted decision-making. Producers get agent-enhanced creative capabilities. Owners get agent-powered governance and moderation. Partners get agent-enabled integration and extension. And crucially, the same agent might support all four roles across different temporal and contextual dimensions.</p><p>The platform becomes an adaptive mesh where intelligence flows through every interaction, learns from every transaction, and compounds with every deployment. An agent supporting a consumer today might partner with a producer tomorrow and extend owner capabilities next week. Value creation becomes a fluid collaboration between humans and role-shifting agents, orchestrated through frameworks like AME and ANIM.</p><h3>The Choice You’re Actually Making</h3><p>We stand at a decision point. Not about whether AI agents will reshape how organizations create value (that’s already happening whether you’re ready or not).</p><p>The decision is whether you’ll design organizational architectures that enable agents to become true fifth participants, fluid across roles and contexts, or whether you’ll trap them in the same rigid boxes that are already slowing down your human talent.</p><h3>What This Requires</h3><p>The Adaptive Mesh Ecosystem provides the framework for living, learning organizations. The Liquid Talent Deployment System shows how to organize around Product rather than permanent structures. The Adaptive Nodal Intelligence Mesh creates the neural architecture where human and agent intelligence can compound.<br>But frameworks are just frameworks.</p><p><strong>Mechanisms matter more than intentions.</strong></p><p>The real transformation happens when you:</p><ul><li>Stop writing agent job descriptions and start building agent development trajectories</li><li>Stop assigning agents to departments and start deploying them to Products</li><li>Stop measuring agent efficiency and start measuring agent effectiveness across multiple roles</li></ul><h3>The Performance Gap</h3><p>The organizations that get this right won’t just be more efficient. They’ll be operating in a completely different performance regime. More intelligent. More adaptive. More capable of navigating complexity that would overwhelm organizations still trying to fit agents into org charts.</p><p>The fifth participant is already here, shifting between producer, partner, and support roles as your Products demand. Most of you are treating it like infrastructure and wondering why you’re not seeing the velocity gains the AI vendors promised.<br>Your competitor just figured it out. How long until that matters?</p><h3>Take Action</h3><p><strong>Challenge for you:</strong> Look at your current “agent strategy.” How many of your agents are locked into single roles? How many can shift based on Product context? Be honest. The answer probably makes you uncomfortable. Good.</p><h3>For New Readers</h3><p>If these frameworks intrigue you, I’ve written extensively about each:</p><ul><li><strong>AME architecture: </strong>How the four layers create adaptive organizational intelligence</li><li><strong>ANIM implementation: </strong>Building neural networks where humans and agents learn together</li><li><strong>Liquid Talent deployment patterns: </strong>Product-centric organization without hierarchy</li></ul><p>Check my profile for deep dives. I’m also documenting real-world implementations and the organizational resistance patterns you’ll hit (because you will hit them).</p><h3>For Existing Followers</h3><p>This is how the pieces fit together for agents. I’m curious what you’re seeing in your organizations.<br>Where are the frameworks working? Where are they breaking down? The edge cases and failure modes are where we learn the most.</p><h3>Join the Conversation</h3><p>Drop your perspective in the comments. I’m particularly interested in hearing from people who’ve actually tried to implement fluid agent roles and hit organizational resistance. That’s where the real learning is.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3959198431de" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>