<?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 Hardik Goel on Medium]]></title>
        <description><![CDATA[Stories by Hardik Goel on Medium]]></description>
        <link>https://medium.com/@hardik.goel214?source=rss-1f6541a8c84a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*-ZSwJfuQavYfMy73.jpg</url>
            <title>Stories by Hardik Goel on Medium</title>
            <link>https://medium.com/@hardik.goel214?source=rss-1f6541a8c84a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:30:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@hardik.goel214/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[Timer, Judge, Schedule. Three Triggers, One Agent , and a Month of Forcing Everything Into Cron.]]></title>
            <link>https://medium.com/@hardik.goel214/timer-judge-schedule-three-triggers-one-agent-and-a-month-of-forcing-everything-into-cron-6343b3d29914?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/6343b3d29914</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[trigger]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-11-13T21:31:01.641Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>A hands-on guide to Hermes Agent’s /loop, /goal, and cron went round this week. It’s the most practical loop-engineering write-up I’ve seen, and its best insight isn’t a feature. It’s a taxonomy.</blockquote><p>If you’ve read my graph-engineering piece, you’ll remember the argument: nothing died, everything got wrapped, and a loop is a one-node graph. That was the theory. This week internet published the practice, three days into using Hermes Agent’s brand-new /loop command (shipped August 13) alongside the older /goal and cron, and the useful part isn’t the commands. It’s the confession running underneath them: a month of forcing everything into cron “because that’s what I knew from Unix,” and the slow realization that cron was the wrong shape for half of it.</p><p>So, in my usual manner: agree with the guide, argue with a word. The word is <em>automate</em>. The title promises to automate anything. What the guide actually teaches is something more useful and more precise: there are three <em>shapes</em> of unattended work, each with a different trigger, and most automation mistakes are using the wrong one.</p><h3>Three shapes, three triggers</h3><p>The 30-second version, which the author rightly puts first. Want the agent to poll or watch something <em>during this session</em>? /loop. Want it to keep working on one defined objective <em>until it’s done</em>? /goal. Want work to run on a schedule, unattended, surviving restarts? cron. Each uses a different trigger, timer, judge, schedule; lives in a different place, session, session, its own process; and ends differently, stop condition, goal achieved, you delete the job.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*HcZl0Kj-xxPGB5yz.gif" /></figure><p>That table is the whole article, and it’s the same architecture lesson as the last five things I’ve written: the trigger is code (a timer, a schedule, an exit code), the judgment is model, and the mistakes all come from putting the model where the trigger should be, or a schedule where a session should be.</p><h3>/loop: the tool for “watch this while I work”</h3><p>/loop 2m check the deploy repeats a prompt on a timer inside your live session until a stop condition fires. Two cadence modes, and the author’s default is the smart one: <em>self-paced</em>, where Hermes starts at a minute, backs off to 2, 4, 8, up to 15 minutes if nothing changes, and snaps back to a minute the moment something does. Exponential backoff for attention. You stop guessing whether two minutes is too often.</p><p>Two details worth pinning. A wakeup is an ordinary user-role turn, no system-prompt mutation, so prompt caching stays intact and the loop <em>remembers what it already reported</em> — which is exactly why the author’s “scan X for trending posts every 2 hours” moved from cron (fresh session, duplicates everywhere) to /loop (one session, no duplicates) and isn’t going back. And loop state survives /resume and context compression but <em>dies with the session</em>, which is Mistake 2 in the author’s own list: /loop 30m watch the queue, close laptop, sleep, wake to nothing. Overnight is cron. Period.</p><h3>/goal: the tool for “keep going until it’s done,” and the trap inside it</h3><p>/goal fix every failing test in tests/hermes_cli/ runs turn after turn; a small judge model checks after each turn whether the goal is met and either marks it done, sends a continuation prompt, or returns <em>wait</em> and parks the loop when it’s blocked on CI or a build (no turn burned, no judge call, resumes when the process exits). Twenty-turn budget by default, then it pauses for you to /goal resume in measured chunks. The judge is deliberately conservative and costs about 200 tokens a turn, so a cheap fast model is right for it.</p><p>Here’s the trap, and it’s the author’s Mistake 3, and it’s the checkability line from my Andrew Ho piece wearing a slash command. /goal make the auth flow work: the judge marked it done on turn three because the agent <em>said</em> it was working; tested manually, it wasn’t. The judge reads prose. Vague target, vague judgment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*pwpgDsz4AMiWRT2i.gif" /></figure><p>The fix is a <em>completion contract</em>, verify: pytest tests/auth passes, boundaries: only touch services/auth, stop when: a schema migration is required, which you can have Hermes draft from a one-liner, plus a <em>quality gate</em>, a shell command that must exit 0 before the judge even runs. Gates pass, normal judging; any gate fails, its output becomes the continuation prompt. An LLM judge stacked on a deterministic check: the exact layering Airbnb described in its eval write-up, for the exact reason, prose lies and exit codes don’t. Give the judge an answer key.</p><h3>Cron: the tool for “run this whether my laptop is open or not,” and where the money goes</h3><p>Cron lives outside all sessions, survives restarts, and every tick is a <em>fresh</em> agent session with no prior context, which means, the author stresses, prompts must be self-contained: not “check on that server issue” but the IP, the user, the systemctl command, and the URL that should return 200. Assume the agent knows nothing except the prompt.</p><p>Then the part I’d have every founder read twice: the three cost tiers. Tier 1, full LLM every tick. Tier 2, a <em>wakeAgent gate</em>: a script runs first, and if its last line is {&quot;wakeAgent&quot;: false} the model is never called, zero tokens on empty ticks. Tier 3, no_agent: no model at all, the script’s stdout is delivered verbatim, empty stdout means silence, non-zero exit means an error alert, so a broken watchdog can’t fail silently.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*ym93NskR8pE7f3Fm.gif" /></figure><p>Mistake 4 is the number: an email check every two minutes, 720 LLM calls a day, 700 of them returning “no new emails,” money on fire, until the gate. Read it as a rule and it’s the reducer article’s rule again: <em>code decides whether to wake the model; the model doesn’t get to decide whether it was worth waking.</em> Every polling job that runs more than once an hour wants a gate in front of it. There’s also [SILENT] suppression for jobs that should only speak when something’s wrong, and context_from to chain isolated cron jobs, collector at 7, triage at 7:30, brief at 8, so each gets the previous one’s output.</p><h3>How they compose, which is where it gets good</h3><p>Real setups use two or three. /loop + /goal in one session: the goal owns the session while driving; when the judge parks it on a <em>wait</em>, the loop takes the idle time and heartbeats the deploy dashboard every two minutes without burning goal turns; when the deploy finishes, the goal unparks and takes the session back. /goal + cron: cron fires a session on Monday at 8 whose prompt <em>is</em> a /goal with a contract, so scheduled long-horizon work gets judge-verified completion. /loop + cron don’t interact, different scopes, and if you want loop behaviour overnight, that’s cron with a wakeAgent gate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*KDocJIAEkqU6OSig.png" /></figure><p>The full stack, in the author’s phrasing that I’d frame: cron owns the trigger, /goal owns the deliverable, /loop owns the ongoing attention. Cron fires at 7 on a weekday, /goal drives the morning brief to a verified finish, then a /loop watches the inbox every ten minutes until you close the session at six, and tomorrow it all fires again. Zero dollars on the idle ticks.</p><h3>The five mistakes, and why I trust the guide</h3><p>The best section is the author’s own error log, and I’ve kept it whole because self-reported mistakes are the most reliable documentation there is.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*jMdoR0gVYp4KFkzX.png" /></figure><p>Cron for in-session CI polling (fresh session, no context, use /loop). /loop for overnight work (dies with the session, use cron). /goal without verify criteria (the judge reads prose, give it a contract or a gate). Cron polling without a gate (720 calls, 700 empty). And /loop without a stop condition (went to lunch, the tick-100 backstop saved him; always --times N or --until X, and leave max_ticks at 100).</p><h3>Where I’d push, gently</h3><p>Two things I’d add. First, the composition rules are elegant, but they’re also <em>implicit state</em>: which of the three currently owns the session is something you have to know, and I’d want that surfaced as loudly as the progress bar in the tab header, because “why did my loop stop firing” will otherwise be a support ticket. Second, the --until judge and the /goal judge are the same auxiliary model, and everything from the Airbnb piece applies: calibrate it on a small golden set of <em>your</em> goals, including ones that should <em>not</em> pass, before you trust it overnight. A conservative judge is a good default; a measured judge is a good system.</p><p>I’ll also admit a proprietary interest. I’ve spent this year building an orchestration dashboard whose whole premise is bounded workflows with gate-integrated approvals, and reading this guide was like reading someone independently arrive at the same three shapes from the command line up. Timer, judge, schedule. It’s a good sign when the practitioners and the architecture diagrams converge; it means the shape is real and the names will keep changing.</p><p>Pick by shape, not by habit. Cron owns the trigger, the goal owns the deliverable, the loop owns the attention, and code, not the model, decides when to wake anyone up.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems, a suspicious number of which began life as a cron job someone should have made a loop. I write about AI’s mechanics, its hype cycles, and the boring scheduling decisions that decide whether the clever bit runs at 3 a.m. Subscribe; it’s a wakeAgent gate that always returns true.</em></p><p><strong>Sources / further reading:</strong> YanXbt (@IBuzovskyi), “Hermes Agent: automate anything with /loop, /goal, and cron. Full guide” (X, Aug 16, 2026), the source for the decision tree, the command syntax and configuration, the three cron cost tiers, the composition rules, the surface-by-surface notes, and the five mistakes; summarized with commentary. Nous Research’s Hermes Agent documentation for loops, goals, and cron (hermes-agent.nousresearch.com/docs); /loop shipped in Wingtips #46 (Aug 13, 2026). My earlier pieces on graph vs loop engineering, the reducer between workers and synthesis, and Airbnb’s eval layering, which this connects. The “code decides when to wake the model” and calibration additions are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6343b3d29914" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Agent Can Think. Can It Remember? Wrong Question. Ask Who Gets to Remember What.]]></title>
            <link>https://medium.com/@hardik.goel214/your-agent-can-think-can-it-remember-wrong-question-ask-who-gets-to-remember-what-4404a2028318?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/4404a2028318</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-11-06T21:31:01.731Z</atom:updated>
            <content:encoded><![CDATA[<p>A clean buyer’s guide to agent memory went round this week: three questions, ten systems, one honest personal choice.</p><p>Two memory pieces in as many weeks, and they’re complementary. An essay, which I wrote about last time, was the <em>physics</em>: storage is solved, retrieval is the job, forgetting is a feature. This one, from Hermes Release Watch, is the <em>shopping guide</em>: how to pick a memory layer without reading every architecture paper first. It’s practical, unhyped, and it opens with the sentence every agent user has lived: your agent can solve a hard problem today and still make you explain half of it again tomorrow. That’s the gap between intelligence and memory, and no amount of “better model” or “bigger window” closes it on its own.</p><h3>The four-tier mental model</h3><p>The piece’s framing is homely and correct. Your agent has a <em>desk</em>: what’s in front of it right now, the context window. It may have a <em>notebook</em>: the small, durable things, preferences, important facts, decisions, lessons. There’s an <em>archive</em>: the big body of past material it can <em>search</em> when needed, not reread. And once you run several agents, there’s a <em>shared workspace</em>, selected knowledge deliberately available to more than one of them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*2_a2nt_9jmFdkbXy.gif" /></figure><p>The line I’d frame: a transcript stores what happened; a useful memory system helps the agent find the part of what happened that matters now. That’s “retrieval is the job” in a cardigan. And the last tier is the tell: the moment you have five agents instead of one, memory stops being a nice extra and becomes part of the architecture, because now the question isn’t only <em>can it remember</em> but <em>which of them</em> remembers <em>what</em>.</p><h3>Don’t pick a system. Answer three questions.</h3><p>There are, the piece notes, a lot of projects using the word “memory” now. Its filter is three questions, and I’d endorse them as the entire buying process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*DPoqCuvR20I_1jWg.gif" /></figure><p><strong>Where should the memory live?</strong> Local or self-hosted, where you control the layer, the data, the backups, and, a nice point, your agents can help monitor and maintain it; or managed cloud, simpler to start, but you accept the provider’s architecture, data handling, limits, and service model. Neither is automatically better. Rent-versus-own again, and my bias from last time is on record.</p><p><strong>How does the agent find the right part?</strong> Storing is half the problem; if the agent rereads thousands of old messages to locate one fact, you have persistent memory but not an efficient system. The example given is OpenViking’s three context levels: a short L0 abstract for quick filtering, a larger L1 overview for navigation and decisions, and the full L2 material loaded only on demand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*O5bedCj32byuKpht.gif" /></figure><p>The plain-English version the author gives is the one to remember: your agent shouldn’t have to reread the whole filing cabinet to find one folder. Read the smallest thing that answers the question; escalate to the bigger thing only when it doesn’t. That’s the reducer article and the memory article agreeing across a table.</p><p><strong>Who should be able to use the memory?</strong> With one agent, it hardly matters. With five, it’s everything. Some knowledge belongs to one agent, some to one user or project, some should be deliberately shared, and you probably do not want every agent automatically inheriting every personal preference, customer record, or internal state.</p><h3>Where I’d argue: the third question isn’t a memory question</h3><p>Here’s my quibble, and it’s meant to sharpen the piece, not dent it. The author frames question three as a memory-design decision. I’d frame it as an <em>access-control</em> decision that happens to be about memory. “Can I control who gets to remember what?” is the same question as “can I control who gets to read what?” and the answer needs the same machinery: scopes, ownership, and a default of <em>not shared</em> until someone says otherwise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*OcGqVCHb_VFrIhWg.png" /></figure><p>Regular readers will recognize the shape from my Grok Bot piece: five named bots on one machine, one sign-in becoming everyone’s, and the vendor’s own documentation warning not to use separate bots as a security boundary. Shared memory without scopes is the same failure wearing a notebook. The monitoring agent that discovered a customer’s billing dispute today should not be the reason the research agent “knows” that customer’s card details tomorrow. And the OWASP entry for memory poisoning applies with extra force in a shared workspace: one poisoned note, trusted by five agents, is a supply-chain attack you built yourself.</p><p>So my amendment to the three questions: ask the third one <em>first</em>, and ask it as a security question. It’s the one that determines whether a shared workspace is an asset or a blast radius, and it’s the one the single-agent tutorials never mention because they never had to.</p><h3>The ten, and the honest personal choice</h3><p>The piece narrows the field to ten systems and gives each a line, which I’ve reproduced as a card, in its own words and no more, because that restraint is the point: it isn’t a ranking, they’re different answers to different problems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*CDgqfTF0LH0LN7YW.png" /></figure><p>The most useful paragraph is the author’s own switch: from Mem0 as an external layer to OpenViking as the Hermes setup grew into a network of specialized agents, because they wanted the memory local, inspectable, and maintainable by the agents themselves, and the tiered retrieval fit. And then the sentence that makes me trust the whole thing: this doesn’t make OpenViking the universal winner; it’s an example of choosing memory based on the architecture you’re building. Two commenters immediately posted <em>their</em> different choices, one on the default Hermes memory, one on Hindsight plus Obsidian, one on a 426-token-prompt approach, which is exactly what a healthy tool ecosystem looks like and exactly why “which is best” is the wrong question.</p><p>Also worth pinning: strong memory doesn’t have to mean another subscription. Most of the ten have an open-source or self-hosted path; you may still pay for embeddings, extraction, or storage, but not a recurring memory-service bill. That’s the same “own the vault” argument from last week, said gently.</p><h3>The rule to keep</h3><p>Don’t choose memory because one project has the longest feature list. Answer where, how, and who, and the list gets small. A smarter model determines what your agent can do right now. Memory determines what useful parts of that work the system gets to keep. And once several agents build on yesterday’s work, memory isn’t bolted on later. It’s part of the system.</p><p>To which I’d add one line, from the last three articles running in my head at once: what gets kept is a retrieval problem, what gets shared is a permissions problem, and if you solve the first without the second, you’ve built a very efficient way to leak.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data systems, a fair number of them multi-tenant, which is why “who can read this” is the first question I ask of anything that persists, agents included. I write about AI’s mechanics, its hype cycles, and the boring permissions layer that decides whether the clever bit is safe. Subscribe; I’ll scope it to you.</em></p><p><strong>Sources / further reading:</strong> Hermes Release Watch (@HermesWatcher), “Your AI Agent Can Think. Can It Remember?” (X, Aug 16, 2026), the source of the desk/notebook/archive framing, the three questions, the ten-system list and their one-line descriptions, and the Mem0 → OpenViking account; summarized with commentary, and the ten descriptors are theirs verbatim in spirit and no longer. OpenViking’s L0/L1/L2 context levels as described there. My earlier pieces on memory engineering (retrieval is the job) and on Grok Bot’s shared computer (bots are not a security boundary), which this connects. OWASP’s memory-and-context-poisoning entry. The “third question is access control” framing is mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4404a2028318" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Tesla Tore Out the Model S Line in 46 Days to Build Robots.]]></title>
            <link>https://medium.com/@hardik.goel214/tesla-tore-out-the-model-s-line-in-46-days-to-build-robots-25e269ccfca6?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/25e269ccfca6</guid>
            <category><![CDATA[humanoid-robot]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tesla]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[elon-musk]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 21:36:01 GMT</pubDate>
            <atom:updated>2026-10-30T21:36:01.592Z</atom:updated>
            <content:encoded><![CDATA[<h3>Tesla Tore Out the Model S Line in 46 Days to Build Robots. The Hard Part Isn’t the Body. It’s That There’s No Internet of Touch.</h3><blockquote>A good primer on Physical AI went round this week under an Optimus headline. The headline is about a factory. The essay is about a data problem, and it’s the data problem that decides everything.</blockquote><p>The facts first, because they’re real and dramatic. On the Q4 earnings call in January, Musk gave the Model S and Model X an “honorable discharge.” The last ones rolled off in Fremont in May. Tesla then posted a time-lapse of the original S/X assembly line, the one that made it a carmaker in 2012, being torn out in 46 days, concrete trenches and all, to make room for a line designed for up to a million Optimus humanoids a year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*n95hc0enLJUJc4RZ.gif" /></figure><p>Two footnotes the reposts skip. That million is <em>design capacity</em>, not a shipped number; Tesla’s own recent updates have pushed the timeline back, each Gen 3 robot has around 10,000 unique parts on an entirely new supply chain, and Musk himself called initial output “extremely slow” and the near-term pace impossible to predict. And “Physical AI” and “humanoid robot” are not the same thing, which is the essay’s best point and the one I want to build on.</p><p>So: agree with the primer, argue with a word. The word is <em>humanoid</em>. It’s the least important part of the story and it’s the part every headline leads with.</p><h3>The body is not the brain</h3><p>The essay’s cleanest distinction: a humanoid is a <em>form</em>, a machine shaped like us that walks, grips, climbs stairs. Physical AI is the <em>intelligence</em>, a system that perceives its surroundings, understands context, decides, acts, and, the operative word, adapts when the situation changes. A body doesn’t imply the brain. Most of the impressive demos, the dancing, the coffee-making, are imitation learning inside a rehearsed scenario: the robot reproduces a sequence, and if you move the mug it does not necessarily know what to do next. And the brain doesn’t need the body: Physical AI can live in an industrial arm or a car.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*dsZnuVpzkRKhanib.gif" /></figure><p>That reframe puts a sharp question under every viral clip: not “can it move like us?” — that part is increasingly solved — but “does it change its plan when the world changes?” Regular readers will hear the checkability line from my Andrew Ho piece. Physics <em>has</em> an answer key; the cup either falls or it doesn’t. What it doesn’t have is a <em>dataset</em>.</p><h3>The actual bottleneck: what tokenizes, and how much of it exists</h3><p>Here’s where the essay earns its length, and where I’d put the whole story. LLMs got good because the internet is an enormous, free, pre-tokenized corpus of text. Vision followed because images and video could be turned into tokens too, so a robot inherits, from pretraining, that water is liquid and sand is granular without relearning it from scratch. That’s why vision-language-action models exist and work at all: they bolt action heads onto the sensory channel we already had data for.</p><p>Now go down the sensory list and ask two questions of each: does it tokenize, and how much data exists?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*tGhjiGvqsQ9DehNi.gif" /></figure><p>Depth and LiDAR: tokenizable, but datasets that pair images, point clouds, and <em>correct robot actions</em> at the same time are rare. Sound: tokenizable, but almost never tied to the physical outcome of an action, so the click that confirms a latch closed and the motor whine that means “overload” carry information no model has been trained to hear. Touch and force: the channel that actually decides whether a grip holds, whether the cloth is folding, whether “carefully” was honoured, and there is essentially no corpus for it, because too many physical parameters are lost when you try to describe it in words. Proprioception, the robot’s sense of its own joints, loads, and balance: per-robot telemetry, invisible from outside, and the thing that stops a humanoid from falling over while lifting a box a camera can see perfectly well. Smell and taste: lab samples.</p><p>That’s the sentence I’d frame from all of this: LLMs learned from the internet, and there is no internet of touch. Every impressive robot video you’ve seen this year is running on the top two rows of that chart. Everything a universal robot needs to be <em>useful</em> rather than <em>impressive</em> lives in the bottom four.</p><h3>Why the architecture looks like my last three articles</h3><p>The essay notes that nobody has yet cracked the giant end-to-end “universal brain” that takes video, text, audio, touch, and joint state and emits actions, and that the practical path is modular: one model for vision, one for motion, one for navigation, a language model as high-level planner that turns “pick up the red cup and put it on the table” into locate, position, identify, plan, grip, execute. Improve the motion model by 10% and swap it in without rebuilding everything.</p><p>If that sounds familiar it’s because it’s the same shape as the graph-engineering piece and the reducer piece: judgment nodes, coded edges, and a planner on top. Robotics got here first out of necessity — a real-time balance controller cannot wait for a cloud round-trip, so low-level control lives on the robot while slow reasoning can live in the cloud where a few seconds’ delay is fine. Physical AI is graph engineering with a latency budget you can’t negotiate, because gravity doesn’t retry.</p><h3>From action to prediction</h3><p>The essay’s forward look is the World-Action Model: today’s VLA is roughly perception plus instruction equals action; the next step is a system that models what happens <em>after</em> the action, an internal simulation of consequences. If I push it, where does it go? If I squeeze harder, how does it deform? If I place the load there, do I stay upright?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*OHVMZGjfW512JktO.gif" /></figure><p>Humans run this simulation continuously and never notice. For robots it’s one of the hardest open problems, and today generating those scenarios is too slow for real-time control. Which is a polite way of saying: the thing that would make Optimus <em>general</em> is the thing that doesn’t run fast enough yet, and the thing that makes it <em>impressive</em> is the thing that already does. Judge the demos accordingly.</p><h3>The one place I’ll defend the humanoid form</h3><p>Having argued the body is the least important part, let me give it its due, because the essay’s own case for it is the right one and it’s not “it looks like us.” It’s that the world was built for this body. Doors, stairs, corridors, tools, levers, work surfaces, machinery: all designed for humans. A humanoid can theoretically use all of it without expensive changes to the environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*gdtaCadRvDvq9Zl8.png" /></figure><p>In software terms, the humanoid form is a <em>compatibility layer</em>, and if that phrase sounds familiar it’s because I used the same logic about browser agents two weeks ago: they log in as you instead of calling APIs precisely because the world already has a UI built for a person. Optimus is the physical version of that bet. The value isn’t the resemblance. It’s the zero-integration cost with a world we already built. And, exactly as with the browser agents, the compatibility layer is the <em>cheap</em> part; the intelligence you pour into it is the expensive, unsolved, and interesting part.</p><h3>What has to be true before “in every house”</h3><p>The comment under the original post said “soon they’ll be in every house.” Here’s the sober checklist the essay implies. The cost has to fall, and it isn’t only actuators: it’s compute, training, maintenance, and enormous engineering; humans are still cheaper and more general for most tasks, though technology costs do collapse at scale, as they did for computers, phones, and cameras. The sensing gap has to close, particularly touch, force, and proprioception, and that means new datasets nobody has yet collected at scale. And the platform model has to arrive: one universal body, different skill packs for a hotel, a factory, a kitchen, so capability is mostly software, which is where the real economics of humanoids would come from.</p><p>The turning point, the essay says and I agree, is a machine that reliably does a broad range of real tasks <em>without constant tuning and supervision</em>. Not the dance. The Tuesday morning where the box fell in an unexpected place.</p><p>Forty-six days to tear out a car line. Considerably longer to build an internet of touch. Tesla has bet a factory that the second is coming. I’d bet the intelligence arrives before the body stops looking like a demo, and that when it does, we’ll spend far less time watching robots walk and far more asking the question this essay quietly puts at the centre: not what it can do, but whether it knows what happens next.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems, most of them for a physical world of shelves, warehouses, and trucks that never once handed me a clean dataset. I write about AI’s mechanics, its hype cycles, and the difference between a demo and a Tuesday. Subscribe; unlike Optimus, delivery is immediate.</em></p><p><strong>Sources / further reading:</strong> The Physical AI primer by Mikadzyki (@Mikadzyki_NFT), “Elon Musk Is Cutting Tesla Car Production for Optimus. Why Physical AI Is the Next Revolution” (X, Aug 17, 2026); the humanoid-vs-Physical-AI distinction, the sensory-channel walk-through, the modular-architecture argument, and the World-Action Model framing are his, summarized with commentary. Tesla facts: CNBC on the Q4 2025 earnings call and the S/X wind-down (Jan 28, 2026); Yahoo Finance and others on the 46-day Fremont teardown and Optimus line conversion (July 2026); Tesla’s own subsequent timeline updates noting “extremely slow” initial output and roughly 10,000 unique parts per Gen 3 unit; the 1M/yr figure is design capacity, per those reports. Related background: vision-language-action models and world-action models as described in the primer; NVIDIA Jetson Thor for on-robot compute. The “no internet of touch,” “compatibility layer,” and “graph engineering with a latency budget” framings are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25e269ccfca6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Agent’s “Second Brain” Gets Dumber as It Grows. That’s Not a Metaphor. It’s a Benchmark.]]></title>
            <link>https://medium.com/@hardik.goel214/your-agents-second-brain-gets-dumber-as-it-grows-that-s-not-a-metaphor-it-s-a-benchmark-8ec24cb55c1d?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ec24cb55c1d</guid>
            <category><![CDATA[gpt-4]]></category>
            <category><![CDATA[metaphor]]></category>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[artifical-intellegence]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-10-23T21:31:01.714Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>The best thing I’ve read on agent memory this year says the quiet part first: storage is solved, retrieval is the job, and a memory that forgets on purpose beats one that remembers everything.</blockquote><p>This title, as the author of a book about digital forgetting (if not yet read, go ahead and read it <a href="https://www.amazon.in/Forgotten-Design-Epidemic-Digital-Dementia/dp/9367078056/ref=sr_1_1?crid=AD2H7ITV0OE3&amp;dib=eyJ2IjoiMSJ9.qVchI-3J36KGv0kfTFyjFw.Zp1i9uqqVPdhMTrFosjfb5o7NpLfMz4Cq06VvajaIhU&amp;dib_tag=se&amp;keywords=forgotten+by+design+hardik+goel+book+paperback&amp;qid=1788961169&amp;sprefix=forgotten+by+design+hardik+goelbook+paperback%2Caps%2C254&amp;sr=8-1">here</a>), I found equal parts vindicating and mildly alarming.</p><p>A long, dense, unusually well-cited piece went round this month, the capstone to a loop → graph → memory series, and since I’ve now written the loop-and-graph piece from my side, I owe it the third. Its subject: why your agent boots from zero every session, why “save more” is the fix everyone reaches for, and why that fix quietly makes things worse.</p><p>The number that carries the argument comes from LongMemEval, the benchmark built to pin long-term memory down. Hand GPT-4o only the sessions that actually contain the answer and it scores 0.924. Hand it the full ~115,000-token history, everything it ever “remembered,” and it drops to 0.640. A 30% relative collapse produced by nothing but adding more memory. Weaker models don’t dip, they fall off a cliff: Llama 3.1 70B goes 0.848 → 0.286.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*vqR_YrMImY8UStk1.gif" /></figure><p>Write it on the wall, as the piece says. Then let me do my usual thing: agree with the substance, and argue with a word. The word this time is <em>memory</em>.</p><h3>It isn’t memory. It’s a search engine with a forgetting policy.</h3><p>The industry ships “memory” as a toggle: a settings pane, a checkbox, a black box that quietly gets smarter. The reframe is that memory is an engineering discipline with fifty years of settled science under it, and the model is small enough to hold in your head: three types, four operations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*xRPSwoEieEFaJ9oZ.gif" /></figure><p>The three long-term types come straight from the CoALA agent-architecture taxonomy and are older than transformers: episodic (what happened, dated, in order), semantic (what’s true, stripped of when you learned it), procedural (how to act, the checklist, the code). The split episodic from semantic in 1972; Cohen and Squire added “knowing that” versus “knowing how” in 1980. The fourth type, working memory, is your context window, small, and, in the line I’d frame, attention is a budget you are always overspending. The four operations map onto how memory has been studied since 1963: write, retrieve, decay, consolidate.</p><p>Here’s why I’d rename the whole thing. Every one of the four operations, and every number in the piece, is about <em>finding</em>, not <em>keeping</em>. Keeping is a filesystem. What we’re actually building is a retrieval system whose defining feature is what it declines to show. Call it memory and people build warehouses. Call it a search engine with a forgetting policy and people build the right thing: an index, a ranker, and a bouncer.</p><h3>Write like a bouncer</h3><p>The write itself is trivial; deciding what earns a slot is the entire skill, because every line you write is a line something else now has to retrieve <em>around</em>. The mechanic anyone borrows is from Generative Agents: score each candidate memory for poignancy, 1 to 10, at write time, and don’t write below a bar. Brushing teeth is a 2. A breakup is a 10. In a plaintext vault it’s one line of frontmatter.</p><p>The trap made in the article names the one I’ve spent a book on. Christian Tietze called it the Collector’s Fallacy in 2013: knowing <em>about</em> something isn’t knowing it, and filing things away doesn’t lead anywhere. Andy Matuschak is blunter, that hoarding notes feels like progress and we systematically overrate it. Every save is a Skinner-box pellet. The pile that feels like an asset is a liability with a nice UI, and, the sentence to underline, the goal of the write operation is to write <em>less</em>.</p><p>I’ll add the personal note since the piece demands specificity: I wrote <a href="https://www.amazon.in/Forgotten-Design-Epidemic-Digital-Dementia/dp/9367078056/ref=sr_1_1?crid=AD2H7ITV0OE3&amp;dib=eyJ2IjoiMSJ9.qVchI-3J36KGv0kfTFyjFw.Zp1i9uqqVPdhMTrFosjfb5o7NpLfMz4Cq06VvajaIhU&amp;dib_tag=se&amp;keywords=forgotten+by+design+hardik+goel+book+paperback&amp;qid=1788961169&amp;sprefix=forgotten+by+design+hardik+goelbook+paperback%2Caps%2C254&amp;sr=8-1"><em>Forgotten by Design</em></a> about what outsourcing memory does to human minds, and I’m currently building a personal-capture product. Reading a rigorous case that the healthiest memory system is one that <em>declines to remember</em> most things was, professionally, a very pointed mirror. Forgetting isn’t the failure mode. Indiscriminate remembering is.</p><h3>Retrieve: presentation beats volume, and the window is mostly fiction</h3><p>Four numbers form the spine, and Roxx sources each. Chroma’s Context Rot study, nearly 200,000 LLM calls across 18 models, found accuracy degrades as input grows <em>even when every added token is relevant</em>; what matters is how information is presented, not how much. Lost in the Middle showed the U-curve, and that past roughly 20 retrieved documents accuracy fell below closed-book. NoLiMa put a number on “effective length,” the point where a model still holds 85% of its short-context score, and it’s a fraction of the advertised window: GPT-4.1 advertises a million tokens, effective is around 16K. And the LongMemEval collapse from the top.</p><p>Cognitive science had it in 1975 as the cue-overload principle: the more memories attached to a retrieval cue, the worse the cue gets at retrieving any one of them. A vault where forty notes match #auth is a vault where #auth retrieves a blur. Your second brain doesn’t get smarter as it grows. It gets foggier.</p><p>The honest caveat, and I like that he includes it: it isn’t true that all extra tokens hurt; random unrelated documents can even <em>help</em>. The killer is near-miss memories, semantically similar, subtly wrong neighbours sitting next to the right answer, and cruelly, the better your retriever, the more of them it drags in. Precision, not recall, is what saves you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*BasIF_DwVimA838p.gif" /></figure><p>The mechanic is small enough to be beautiful: score every candidate by recency, importance, and relevance, equal weights, take the top few, and stop. Two details he flags that people get wrong: the recency decay factor in the original Generative Agents paper is 0.995, not 0.99, and it decays per hour since the memory was last <em>retrieved</em>, not since it was written. Which means a note you keep using stays sharp, and one nobody touches slowly sinks. Retrieval strengthens what it retrieves. Karpicke and Roediger showed the human version in 2008: repeated retrieval produced roughly 80% recall a week later versus 33% for repeated studying. The act of retrieval <em>is</em> the learning.</p><h3>Decay is a feature. Deletion is not the same thing.</h3><p>A memory system where everything is permanently, equally accessible is permanently, equally useless. This is the operation CoALA calls understudied and the one homemade vaults skip, which is why they rot: without a decay policy, new writes silently bury old ones, and last month’s hard-won workaround doesn’t announce that it’s un-findable, it just stops coming back.</p><p>The vault mechanic is a re-ranker, not a shredder: a last_accessed timestamp, a boost/damp multiplier at read time, and a periodic sweep that moves cold notes to an archive folder, still on disk, still greppable, just off the default path. In git, forgetting is a move, and a move is reversible. Nothing is deleted. The fact just gets quieter.</p><h3>The dream loop, and where “own vs rent” gets teeth</h3><p>Consolidation is the offline pass that turns a pile of episodes into clean facts and playbooks: merge duplicates, resolve contradictions, promote the durable, drop the noise. The neuroscience metaphor is honest, hippocampus teaching neocortex overnight, and there’s a real product now: Anthropic’s Dreaming, a Managed Agents research preview, with a partner-reported completion-rate improvement of about six times. Roxx is fair to it and precise about its weight: real feature, real SDK, but the headline number is one customer’s self-report on a gated preview no outsider can rerun, a testimonial with an API in front of it, set against the open Sleep-time Compute paper from Letta and Berkeley whose numbers anyone can reproduce.</p><p>Then the judo, and it’s the best paragraph in the piece. Anthropic justifies Dreaming with a list: the input store is never modified; every change is a versioned audit trail with point-in-time recovery; you can review changes before they land; the outputs are plain-text notes and playbooks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*D9W-VHOBOMxDB_TN.gif" /></figure><p>Read that list again. Never mutate the source, versioned history, review-before-apply, plaintext. That is not a description of a black box. That is a description of a git repository. The consolidation function you’d write yourself is one copy-then-propose-then-apply routine, and the audit trail a managed service sells you is git diff and git checkout --. You aren’t renting the discipline. You’re holding it.</p><p>The rent-versus-own case is then straightforward and unhysterical: consumer chat platforms train and retain by default, exports come as compliance JSON with no import path, formats migrate on the vendor’s schedule, and memory creates lock-in because every conversation makes the service more valuable and harder to leave. None of it villainy, just the physics of a rented service. Plaintext in git inverts every liability: inspectable, diffable, revertible, portable. And, the security note that should be louder, OWASP now names memory poisoning as a threat, a bad memory that persists and gets trusted across sessions; in a black box that corruption is undetectable by design, in git it’s a diff and a one-line revert. You can’t debug a memory you can’t open.</p><p>He’s honest about the limits too, and it’s why I trust the rest: markdown-in-git is private from the vendor, not from everyone (no encryption at rest by default); flat files aren’t a database, past a few thousand notes you need a real retrieval layer; and the backlink graph is less portable than the files it links.</p><h3>Measure hit-rate, not headcount</h3><p>The metric section is where I’d send every “second brain” enthusiast first. Stop measuring how much you’ve stored. Score your system against an oracle ceiling, what it would hit with perfect retrieval, and the gap is your retrieval debt in one number. Watch tokens per query and p95 latency. Vault size is the collector’s metric; it goes <em>up</em> as your system gets worse.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*b7BMB8BzZz7_MRIQ.png" /></figure><p>And distrust the leaderboards, which he documents with relish: a popular memory benchmark whose answer key is wrong 6% of the time and whose LLM judge accepts most wrong-but-adjacent answers; vendor numbers that contradict each other when rerun; a famous GraphRAG “72–83%” that’s a judge win-rate, not accuracy. If you didn’t run it, on your data, don’t trust it, and he explicitly includes the vendor numbers he leaned on. Notice, though, the shape of every first-party win that survives: not “we stored more” but “we kept less in front of the model.” That’s the shape of number to trust.</p><p>Best of all, he tells you when to skip everything: if the history fits in the window and is all signal, paste it in; a few hundred notes need a folder, not a retrieval layer; API and Team tiers already sidestep the consumer-privacy motive; and a literal Ebbinghaus curve bolted on blind can underperform, so add forgetting only after the metric says you need it. Don’t sell a cathedral to someone who needs a shed.</p><h3>The shift</h3><p>The person who “turned on memory” opens a fresh chat next quarter, watches it re-derive the same three facts, and concludes the tech isn’t there yet. The engineer who did the discipline has a plaintext vault they can read, a retrieval path they can measure, a decay policy that keeps the fog off, a consolidation pass they can diff, and a git history that turns drift into a revert. Same model, same window. The difference isn’t what the agent stores. It’s what it can find, and who owns the finding.</p><p>Which is, once more, the pattern of every piece I’ve written this month: storage is a commodity, models are commodities, context windows are getting advertised in millions. The scarce thing is judgment about what to keep in front of the model, and the ability to <em>audit</em> what it thinks it knows. Trust, again, lives in a diff.</p><p>Write like a bouncer. Decay like a re-ranker. Consolidate offline. Measure hit-rate, not headcount. Own the vault, and let the rest decay.</p><p>I wrote a whole book arguing humans should. Nice to see the agents catching up.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data systems and wrote</em> Forgotten by Design <em>about what happens to minds that outsource remembering; it turns out the same rules govern the agents. I write about AI’s mechanics, its hype cycles, and the difference between a memory and a pile. Subscribe, and let this note score at least a 7.</em></p><p><strong>Sources / further reading:</strong> Roxx (@Roxx_0x), “Memory Engineering: retrieval is the job” (X, Aug 6, 2026), including the mtrace reference vault at github.com/Roxx0x/mtrace; the framing, the four operations, the code, and the vendor-vs-git argument are his, and this is a summary with commentary. Papers he cites and I checked: LongMemEval (arXiv:2410.10813); CoALA (arXiv:2309.02427); Generative Agents (arXiv:2304.03442); Lost in the Middle (Liu et al., TACL 2024); NoLiMa (ICML 2025); Chroma’s Context Rot (2025); The Power of Noise (SIGIR 2024); Zep’s temporal knowledge graph (arXiv:2501.13956); Sleep-time Compute (arXiv:2504.13171); Karpicke &amp; Roediger (Science, 2008); Murre &amp; Dros’s Ebbinghaus replication (PLoS ONE, 2015); Watkins &amp; Watkins on cue overload (1975); Tietze on the Collector’s Fallacy (2013). Anthropic’s Dreaming and memory-tool figures are first-party and, as he notes, not independently reproduced. The “search engine with a forgetting policy” rename and the digital-dementia connection are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ec24cb55c1d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[He Put One Function Between 40 Agents and the Model That Reads Them. Cost Fell 86%.]]></title>
            <link>https://medium.com/@hardik.goel214/he-put-one-function-between-40-agents-and-the-model-that-reads-them-cost-fell-86-001ac4376036?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/001ac4376036</guid>
            <category><![CDATA[cost-optimization]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-10-23T21:31:01.702Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>A builder wired forty Claude Haiku workers into one Claude Sonnet report-writer, watched the bill land on the wrong line item, and fixed it with thirty lines of Python</blockquote><p>A post went round this week from a builder running a fan-out research pipeline: forty parallel Claude Haiku workers, each taking a narrow slice of one question, all feeding one Claude Sonnet call that writes the final report. Forty cheap calls, one strong call to close. Sounds efficient. Everyone budgets it that way.</p><p>Then he counted. Raw combined output landing on Sonnet: 41,200 tokens. Fifteen of the forty findings were near-duplicates. Six were malformed. Sonnet’s first job on every run wasn’t reasoning; it was janitorial. In his own words, he’d built a research pipeline and shipped the expensive model a data-cleaning job wearing a research pipeline’s clothes.</p><p>The fix was one deterministic function, no model in it. Same forty workers, same Sonnet: 5,300 tokens in, cost per run from $1.38 to $0.19, latency from 51 seconds to 11, and, the part he didn’t expect, 23 contradictions between workers surfaced that the raw dump had buried.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*91IDLbw11ouWVD-Q.gif" /></figure><p>I want to give the post its due, because it’s the rare viral engineering thread with a runnable function and honest caveats. And then, in my usual manner, I want to argue with the word <em>multi-agent</em>, because the interesting thing here isn’t agentic at all.</p><h3>Everyone budgets the workers. Nobody budgets the reader.</h3><p>The insight in one line: in a fan-out pipeline, the expensive layer is not the fan. It’s the funnel. Forty Haiku calls cost pennies; the model that has to <em>read all forty outputs</em> is billed on input, and input is where the growth is. Add workers for coverage and you’re not scaling the cheap layer, you’re silently scaling the bill of the one call nobody put in the spreadsheet.</p><p>There’s a second cost that isn’t tokens, and the post cites the right paper for it. Liu and colleagues’ <em>Lost in the Middle</em> found that models retrieve information well from the start and end of a long prompt and significantly worse from the middle, even on long-context models. Dump forty unordered findings into one prompt and you haven’t just inflated the bill; you’ve filed most of the evidence in exactly the region the reader is worst at.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*Ye_IFYB1m4C8ntNS.gif" /></figure><p>Which reframes the whole thing. The reducer isn’t primarily saving money. It’s moving the evidence out of the model’s blind spot and telling it, in advance, which claims agree and which collide.</p><h3>The function, and why the surprising part is the important part</h3><p>The reducer does four things, none needing a model: drop entries missing a claim, evidence, or source; group the rest by a normalized claim; keep the highest-confidence version of each group; and, when several workers independently landed on the same claim, <em>say so</em>, append “confirmed by N other workers,” because agreement is a signal the synthesis model should be handed, not something it should have to rediscover by reading forty documents back to back.</p><p>The surprise: grouping doesn’t only dedupe, it <em>exposes disagreement</em>. Two workers asserting contradictory values for the same fact become visible the moment you group instead of concatenate, because a contradiction on page one and page thirty-three of a wall of text is not something a model reliably catches by reading forward.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*B0TYo_dybzU1WZWF.gif" /></figure><p>Twenty-three contradictions across the sample runs, found by a defaultdict. That’s the sentence to keep. Structured reduction isn’t a cost optimization with a correctness bonus; it’s a place to put a check that plain concatenation has <em>no way to run</em>. And the escalation number confirms it: runs kicked to a human for ambiguity dropped from six in fifty to one in fifty, because the ambiguity was being surfaced upstream instead of smeared across the prompt.</p><h3>Where I’d argue with a word: this isn’t a multi-agent lesson</h3><p>Here’s the friendly quibble. The post frames itself as advice for people “scaling a multi-agent pipeline.” But strip the model names and look at the shape: many cheap parallel workers emit key-value-ish records, a deterministic step groups by key and reduces, one expensive step consumes the result. You have reinvented MapReduce. Google published it in 2004. That’s a compliment, not a dunk: the reason MapReduce mattered was precisely that the middle box, the shuffle-and-reduce, was code, and code is where you put the things that must be exact and cheap.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*sivWIku0Nncp9i3k.png" /></figure><p>Faithful readers will recognize the rule from my graph-engineering piece: nodes are model, edges are code. Judgment at the ends of the pipeline, where the workers interpret sources and Sonnet writes; plumbing in the middle, where nothing needs to be smart. What made this pipeline expensive was that a piece of plumbing had been handed to the most expensive node in the graph. The fix wasn’t a smarter model or a cheaper worker. It was taking work <em>away</em> from Sonnet.</p><p>The industry name-cycles I keep writing about make this easy to miss. “Multi-agent,” “orchestration,” “swarm”: the vocabulary implies every problem in the system is an agent problem, so people reach for another model call. Most of the problems in a fan-out system are the same ones a 2004 data engineer had: dedupe, validate, group, sort. Not one of them wants a language model in the loop.</p><h3>The guards, and the bug he told on himself</h3><p>Four guards, all code: malformed entries dropped and logged, not silently absorbed; disagreements flagged and passed to synthesis explicitly; singleton groups passed through with no false confidence added; an empty input returns empty and skips synthesis instead of running Sonnet on nothing. That last one is small and I love it, because “the expensive model ran on an empty list” is a real bill I have seen on a real invoice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*jrrZDbZljXdzhIyu.png" /></figure><p>And then the honest bit, which is why I trust the rest: normalize() is a similarity match on claim text and its false-merge rate is untested. Two genuinely different claims that happen to phrase alike could be collapsed into one, a correctness bug hiding behind a cost win. He says he wants more runs before trusting it on anything he can’t spot-check. Good. I’d add one more caveat he doesn’t mention: “confirmed by three other workers” is only evidence of truth if the workers are <em>independent</em>. If four Haiku workers all read the same upstream blog post, you have one source wearing four hats, and the reducer has just upgraded a rumour to consensus. Group by claim, yes; but also count <em>distinct sources</em> behind each group, and don’t let the confirmation tag fire on a single URL.</p><h3>Steal this</h3><p>If more than one worker feeds one final model, put a reducer between them before you touch the prompt. Measure token count <em>into</em> the synthesis call before you optimize anything else; it’s usually the whole cost story. Treat agreement and disagreement between workers as data the reducer surfaces, not noise the model rediscovers. And keep the boundary clean: dedupe, drop, group, and sort are code problems; reasoning about what’s left is the model’s job. Don’t let one bleed into the other.</p><p>Forty workers were never the expensive part. The reader was. Which is a very 2026 way of learning a very 2004 lesson: the cheap machines can fan out as wide as you like, as long as something boring and exact stands between them and whoever has to make sense of it.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data systems, a good chunk of it writing reducers before anyone called the map step an “agent.” I write about AI’s mechanics, its hype cycles, and the number of times a new vocabulary rediscovers an old pipeline. Subscribe; my synthesis step is one human, and I’d like fewer tokens landing on him.</em></p><p><strong>Sources / further reading:</strong> The original breakdown, posted on X and Telegram by its author, including the reduce_findings() function, the per-run numbers (41,200 → 5,300 tokens; $1.38 → $0.19; 51 s → 11 s; 23 contradictions; 6/50 → 1/50 escalations) and the untested-normalize() caveat; all figures are his, from his sample runs. Nelson Liu et al., “Lost in the Middle: How Language Models Use Long Contexts” (2023; TACL 2024). Jeffrey Dean and Sanjay Ghemawat, “MapReduce: Simplified Data Processing on Large Clusters” (OSDI 2004). My earlier piece on graph engineering for the “edges are code, nodes are model” rule. The independence caveat on the confirmation tag is mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=001ac4376036" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You’re Losing Money. Your Instinct Says Cut. Your Instinct Is Frequently, Precisely Wrong.]]></title>
            <link>https://medium.com/@hardik.goel214/youre-losing-money-your-instinct-says-cut-your-instinct-is-frequently-precisely-wrong-d78b011a34ac?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/d78b011a34ac</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[founders]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-10-16T21:31:01.614Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Part 4 of 4 on IBR’s Mini Strategy Handbook for AI Founders. The last review is a five-minute finance lens that would have saved a lot of businesses I’ve watched shrink themselves to death.</blockquote><p>Here’s the situation the handbook closes on, and if you’ve run anything, you’ve lived it. Revenue has plateaued. Every month closes red. You do the responsible thing: cut marketing, freeze hiring, trim every discretionary line. And the loss… doesn’t move. Or deepens. So you cut again, more responsibly, and it deepens again, and at some point somebody says the word “shut” in a meeting.</p><p><a href="https://www.linkedin.com/in/vivek-bisht101/">Vivek Bisht’s</a> <strong>The Contribution Factor</strong> argues that a lot of these businesses aren’t broken. They’re <em>sub-scale</em>, and every “responsible” cut makes them smaller and therefore worse. The tool that tells the two apart is one subtraction.</p><h3>The formula, and where your next order actually goes</h3><p>Contribution Factor = Price − Variable Cost per Order. It’s what’s left from each sale after paying the costs of making <em>that</em> sale, and it “contributes” toward the fixed costs, rent, salaries, software, until those are covered, after which every further order is profit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*3qW_s8e7pAzKih44.gif" /></figure><p>That’s genuinely the whole thing. But the diagnostic power comes from what the <em>sign and scale</em> of the number tell you to do, and it’s here that most founders’ instincts fail.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*vwy-o5PcS3gsC7uT.gif" /></figure><p>Negative: every sale loses money, scaling makes it worse, fix unit economics before anything else. Positive but low volume: the model works, you’re just too small to cover the cost base, so <em>spend more, not less</em>, and grow into it. Positive and covering fixed costs: you’ve crossed break-even; defend margins, reinvest selectively. The middle row is where a startling number of “failing” businesses actually live, and it’s exactly the row where cutting kills you.</p><h3>The death spiral, and its reversal</h3><p>The handbook’s case is Treya Gifts, an Indian DTC brand selling STEM toys for kids. Loyal customers, strong reviews, eighteen months of plateau and monthly losses, and the founders were doing the textbook thing: slashing marketing, freezing spend, trimming everything variable. Each round of cuts was followed by a proportionate drop in orders, and the losses persisted, because the same fixed costs were now spread across fewer orders. Every cut was defensible. The loop was fatal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*y2Ab648rTmBPkYoC.png" /></figure><p>IBR rebuilt the P&amp;L per order and the contribution factor came back clearly positive, about 33% of revenue. Treya wasn’t selling at a loss. It was sub-scale. So the recommendation was the opposite of every instinct in the room: <em>increase</em> marketing until volume comfortably cleared the fixed-cost barrier, with guardrails, cap cost-per-order below the contribution margin so every incremental order stayed profitable, shift toward channels with clean per-order attribution so you can watch the number live, and freeze fixed costs, no hires, no warehouse, until volume stabilized. Twelve weeks later: orders up 180%, contribution factor steady at 33%, fixed costs up 8%, and the business moved from monthly loss to monthly profit. Not by cutting. By spending more on the one line that was working.</p><h3>Where I’d argue with a word: “responsible”</h3><p>My quibble this time isn’t with the framework, it’s with the vocabulary founders use around it. Cutting costs when you see red ink is described, universally, as “responsible,” and spending more as “reckless.” The contribution lens shows the labels are backwards for a specific, common case. When each order is profitable and you don’t have enough of them, cutting the thing that produces orders is the reckless move dressed as prudence. It <em>feels</em> responsible because it’s painful and it’s what your board expects. Pain is not a signal of correctness. Ask Part 3.</p><p>The handbook flags the misclassification that causes most of this, and it’s the single most useful sentence in the review: <strong>performance marketing is a variable cost, not fixed overhead.</strong> File it under overhead and you will cut it at exactly the moment you should be doubling it.</p><h3>The AI-startup edition, which the handbook leaves for us</h3><p>The Treya case is e-commerce; the handbook is for AI founders. So let me draw the translation, because AI businesses have a variable-cost line that didn’t exist five years ago and that founders routinely misfile.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*we1ksihBDp4GGx3q.png" /></figure><p>Inference tokens per job are variable. So are third-party API calls, human-review minutes, and payment fees. Salaries, cloud baseline, SaaS seats and the model you fine-tuned once are fixed. Which means an AI product can have a <em>negative</em> contribution factor while looking healthy on gross margin slides, if each job burns more tokens than the price covers, and it can be positive-but-sub-scale while looking doomed, if you’ve been treating the ads that bring jobs as overhead. Run the per-job P&amp;L. Put the token bill in the variable column where it belongs. Then read the number and decide, from the table above, not from your stomach.</p><p>One more AI-specific wrinkle: token prices fall. A negative contribution factor today may be positive in two quarters with no change on your side, which is the only case I know where “wait” is a legitimate unit-economics strategy. Note it, date it, and don’t confuse it with the sub-scale case, which waiting only worsens.</p><h3>Closing the series</h3><p>Four reviews, one arc. <a href="https://medium.com/p/34b2393a0ac8/edit"><strong>Launch</strong></a> wide and win by what you cut. <a href="https://medium.com/@hardik.goel214/28d0d012eac1"><strong>Distribute</strong></a> through a specific person with a specific scar, because polish is free now. <a href="https://medium.com/@hardik.goel214/82c90bd36a9b"><strong>Read</strong></a> every signal by what it cost the customer to send. And when the numbers still come back red, <strong>decide</strong> (this one) with one subtraction whether you’re broken or merely small, because the fixes are opposites and the instinct usually picks the wrong one.</p><p>What I like about the handbook, beyond the frameworks, is the through-line I keep meeting everywhere this year: the cheap things, building, polish, attention got cheaper (for more on this, read <a href="https://www.amazon.in/s?k=forgotten+by+design+hardik+goel+book+paperback&amp;crid=AD2H7ITV0OE3&amp;sprefix=forgotten+by+design+hardik+goelbook+paperback%2Caps%2C254&amp;ref=nb_sb_noss">this</a> small book) and the expensive things, judgment about what to cut, a checkable point of view, a customer’s costly commitment, the discipline to read a number instead of a feeling, got scarcer. IBR’s own tagline says it better than I can: judgment is scarce, execution is abundant.</p><p>Which is a strategy handbook’s way of saying what my last four articles said about AI. Nice to know the founders and the models are learning the same lesson.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems, and enough time around P&amp;Ls to have watched good businesses cut themselves into the ground while calling it discipline. This closes my four-part walk through IBR’s handbook. If it was useful, the original is free at ikanabusinessreview.com and worth your subscription; and if you missed Parts 1–3, they’re the launch, the distribution, and the signal you’ll need before this one applies. Subscribe here too, obviously; my contribution factor per reader is positive and I’m merely sub-scale.</em></p><p><strong>Series map:</strong> <a href="https://medium.com/p/34b2393a0ac8/edit">Part 1 · The Inverted Launch</a> → <a href="https://medium.com/@hardik.goel214/28d0d012eac1">Part 2 · The Founder Is the Channel</a> → <a href="https://medium.com/@hardik.goel214/82c90bd36a9b">Part 3 · The Signal Cost Principle</a> → Part 4 · The Contribution Factor (this).</p><p><strong>Source:</strong> Ikana Business Review, <em>Mini Strategy Handbook for AI Founders</em> (August 2026), Review 04, “The Contribution Factor” by Vivek Bisht, including the formula, the decision table, and the Treya Gifts case (+180% orders, 33% contribution factor held, +8% fixed costs, loss to profit in twelve weeks); full case at ikanabusinessreview.com. The “responsible” argument, the AI-startup variable-cost translation, and the falling-token-price wrinkle are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d78b011a34ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Did It Cost the Customer to Send This Signal? Everything Else Is a Footnote.]]></title>
            <link>https://medium.com/@hardik.goel214/what-did-it-cost-the-customer-to-send-this-signal-everything-else-is-a-footnote-82c90bd36a9b?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/82c90bd36a9b</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[founders]]></category>
            <category><![CDATA[customer-experience]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-10-09T21:31:01.478Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Part 3 of 4 on IBR’s Mini Strategy Handbook for AI Founders. You’ve launched, you’ve started distributing, and now numbers are arriving. This is the review about which of them to believe.</blockquote><p>Two founders walk into two investor meetings on the same morning. The first opens with a screenshot: 100,000 on the waitlist, a LinkedIn post with 40,000 views, a launch that “broke the site.” The room nods. The second opens with a sentence: eleven customers fired their existing vendor to switch to us; nine have already renewed. No screenshot. Eleven names.</p><p>One of them has traction. The other has a mirage. And, as <a href="https://www.linkedin.com/in/richa-sati/">Richa Sati’s</a> <strong>The Signal Cost Principle</strong> points out with some sadness, almost everyone in the room is wired to be more impressed by the first one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*JX3pgi4NkKQ33hun.gif" /></figure><p>The whole review, and it’s the longest and most foundational in the handbook, reduces to one question, which the authors themselves say everything else is a footnote to: <strong>what did it cost the customer to send this signal?</strong></p><h3>Signal strength is proportional to signal cost</h3><p>A like costs a thumb-twitch. A purchase costs money. A renewal costs a year of continued trust. A vendor replacement costs budget, politics, and a career-risk bet by whoever championed you internally. These signals were never equal. We just pretend they are, because the cheap ones are so much easier to collect, and because a big number on a slide feels like progress in a way that “nine renewals” doesn’t.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*JnXSCpkfkUC1Tjaz.gif" /></figure><p>The handbook draws this as a Commitment Ladder, and it’s worth internalizing the two instructions that come with it: read signals near the bottom with suspicion; act on signals near the top with confidence. Bigger, louder and prettier are not the same thing as true.</p><p>I’d add the connection to Part 2. The founder channel, done right, generates a <em>lot</em> of bottom-rung signal, views, comments, followers, and it’s intoxicating precisely because it arrives first and in volume. This ladder is the antidote you drink the morning after a post goes well.</p><h3>The famous number was never the number</h3><p>My favourite section: four celebrated companies, and in every case the number they got famous for isn’t the one that convinced them internally. Dropbox’s 75,000-person waitlist from one Hacker News video is the legend; the signal that mattered was people trusting it with real files, inviting colleagues, and returning daily. Slack’s registered users made headlines; the internal bar was teams that exchanged 2,000 messages, of whom 93% were still active later. And Superhuman’s Rahul Vohra used the Sean Ellis test to <em>measure</em> product-market fit instead of arguing about it.\</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*ECdhq9uKSEwgJOFa.gif" /></figure><p>The Superhuman detail is the one to steal. “Do you like the product?” costs nothing to answer, so everyone says yes. “How disappointed would you be if you could no longer use it?” forces the customer to imagine giving something up, so only “very disappointed” counts. The benchmark is 40%. Superhuman scored 22%, and instead of launching, they rebuilt. Slack, tested later, scored 51%. Note the shape of the trick: they made the question cost the customer something, and the truth arrived on its own.</p><h3>The four traps, and their anatomy</h3><p>Every false-traction trap has the same skeleton: a cheap signal, why founders believe it, why it’s weak, and the expensive question to ask instead.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*bppzttkY5a5-1s3u.png" /></figure><p>The Applause Trap: 25,000 views on the demo post; what did anyone actually sacrifice because of it? The Friendly User Trap: friends and advisors say they’d love it; encouragement is free, would you put down a deposit today? The Waitlist Mirage: hundreds joining daily; joining costs four seconds and leaving costs nothing, so how many climbed the ladder, installed, paid, returned? And Pilot Purgatory, the enterprise founder’s special hell: a recognizable logo “testing” you; a pilot carries almost no organizational risk and is not a purchasing decision, so how many pilots <em>replaced a vendor</em>?</p><p>The definition that makes it click: false traction is not fake demand. Dropbox’s 75,000 were real people. It’s <em>misinterpreted</em> demand, a cheap signal read as an expensive one.</p><h3>Where I’d argue with a word, gently</h3><p>The word is “traction.” It’s a physics metaphor: grip, the thing that lets force become motion. But most of what founders call traction is <em>attention</em>, and the review’s own fourth law says it perfectly: marketing creates attention; commitment creates businesses; only the latter compounds. So I’d go one step further than the handbook. Rename the dashboard. Put attention metrics on one page and commitment metrics on another, and never let a number from the first page appear in a fundraising deck without its ladder rung labeled next to it. If your investor can’t tell whether “12,000 users” means “signed up” or “renewed,” you’re not reporting, you’re staging.</p><p>The five laws are worth carrying whole: never optimize a cheap signal; every metric should move customers <em>up</em> the ladder; friction isn’t always the enemy, a deposit or a security review filters intent from curiosity; attention versus commitment as above; and the strongest signal is the one the customer didn’t have to give, an unprompted referral beats any survey. The line I’ll keep, from the review: customers don’t lie, they tell the truth in proportion to what it costs them to tell it.</p><p>Which sets up Part 4 neatly. Suppose you’ve climbed the ladder, real purchases, real renewals, and you’re <em>still</em> losing money every month. The instinct is to cut. The handbook’s last review argues that’s often exactly the wrong move, and it has a formula and a case study to prove it.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems and, on the side, trading Indian equities, where the market delivers the Signal Cost lesson daily: the loud number is rarely the priced-in one. This is Part 3 of 4 on IBR’s handbook. Part 4, The Contribution Factor, is the finance lens that tells you whether to cut or spend, and it’s the one I’d hand every founder staring at red ink. Subscribe; it costs you an email address, which, by this ladder, means I should read it with suspicion.</em></p><p><strong>Series map:</strong> <a href="https://medium.com/p/34b2393a0ac8/edit">Part 1 · The Inverted Launch</a> → <a href="https://medium.com/@hardik.goel214/28d0d012eac1">Part 2 · The Founder Is the Channel </a>→ Part 3 · The Signal Cost Principle (this) → <a href="https://medium.com/@hardik.goel214/d78b011a34ac">Part 4 · The Contribution Factor.</a></p><p><strong>Source:</strong> Ikana Business Review, <em>Mini Strategy Handbook for AI Founders</em> (August 2026), Review 03, “The Signal Cost Principle” by Richa Sati, including the Commitment Ladder, the four traps, and the five laws; full framework at ikanabusinessreview.com. The Dropbox, Slack, and Superhuman figures are as reported there. The “rename the dashboard” suggestion and the trading aside are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82c90bd36a9b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[At Seed Stage You Don’t Have a Brand. You Have a Person. (Bad News for People Who’d Rather Build.)]]></title>
            <link>https://medium.com/@hardik.goel214/at-seed-stage-you-dont-have-a-brand-you-have-a-person-bad-news-for-people-who-d-rather-build-28d0d012eac1?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/28d0d012eac1</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[founders]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[seed-investment]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 21:36:01 GMT</pubDate>
            <atom:updated>2026-10-02T21:36:01.664Z</atom:updated>
            <content:encoded><![CDATA[<h4>Part 2 of 4 on IBR’s Mini Strategy Handbook for AI Founders. This is the review about distribution, and I’m going to write it the way it tells me to: with the scar showing.</h4><p>In <a href="https://medium.com/@hardik.goel214/34b2393a0ac8">Part 1</a> I confessed to shipping eight products this year and being much better at the launching than the cutting. Here’s the second, worse confession, the one that makes Ikana’s Review 02 read like it was addressed to me personally: I can ship anything, and I struggle to distribute it. My own operating rules literally include “outbound is oxygen,” which is the kind of rule you write when you have been, historically, holding your breath.</p><p>So <a href="https://www.linkedin.com/in/richa-sati/">Richa Sati’s</a> <strong>The Founder Is the Channel</strong> is uncomfortable reading, and therefore probably correct. The core claim: at seed stage you don’t have a brand. You have a person. Distribution starts with that person being specific in public, repeatedly, about a problem they’ve actually bled over. Not with ad spend. Not with a logo. With a scar.</p><h3>The polish inversion</h3><p>The mechanism underneath is the sharpest idea in the whole handbook, and it rhymes with things I’ve written about AI deflation. For twenty years, professional-looking content was a moat: a design agency and a copywriter meant you looked credible; without them you looked like a hobby. AI made polished content free, instant, and identical, everyone drawing from the same tools. So polish stopped signaling competence and started signaling <em>nothing</em>, and in some rooms it now signals inauthenticity: if your About page reads like GPT wrote it, and it probably did, the thing that used to build trust quietly erodes it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*4rwgS3S_NsOBfBtf.gif" /></figure><p>Cheap production didn’t democratize credibility. It devalued the currency credibility used to trade in. And what’s still scarce, still expensive to fake? The handbook’s phrase: a specific human with a specific scar, saying something a committee would never approve.</p><p>Read that with the deflation lens from my last few pieces and it clicks: polish is a capability, capabilities got free, and the premium migrated to the one thing you can’t pip install, a person with a checkable track record. Same pattern, different arena.</p><h3>The Swap-the-Logo Test</h3><p>Then comes the tool I’ve already started running on my own drafts, and it stings. Take your last post. Paste it onto a competitor’s account. Does it still make sense? Then you haven’t written anything. You’ve assembled category filler.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*HfXU1BOZBaQSnDPU.gif" /></figure><p>The handbook’s contrast is Pieter Levels, no investors, no PR, just years of building in public with the actual revenue, the actual failures, the actual 3 a.m. decisions, versus the median AI startup account: a rotating cast of “excited to announce,” dashboards nobody outside can verify, enthusiasm calibrated to offend no one. Not wrong. Just interchangeable.</p><p>The distinction I found most useful is that this is a Ben Thompson problem, not a GaryVee problem. “Just be authentic” mostly produces founders <em>performing</em> authenticity, which is its own noise. What actually moves markets is a specific, arguable, falsifiable point of view, delivered consistently, so people know exactly what you think and can check your track record against it. The mechanism isn’t charisma. It’s legibility of belief. “AI will change everything” can’t be checked against anything. “Trust is the last scarce thing in AI, and here’s why” can, and if I’m wrong you’ll get to say so.</p><h3>The SCAR test</h3><p>The review’s operational tool is four questions that are secretly one question, could literally any other founder in your category have written this?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*nMcuIThMm8FqAFJO.png" /></figure><p>Scar, not slogan: “we democratize AI” versus “I hand-labeled data for four months and wanted to set myself on fire.” Contrarian, not consensus: would a smart peer have real reason to disagree? Artifact, not announcement: show the eval, the ticket, the graph of what broke; announcements decay in hours, artifacts get cited for months. Repeated, not one-off: Nomad List wasn’t one good post, it was years of them.</p><p>The Stripe example lands the point: the Collisons wrote deeply technical posts that only made sense if you’d hit the exact walls they described. Nobody hires an agency to fake that level of specificity, because faking it costs more than having the scar. Hence the line I’d frame: the scar is the moat, the one thing a better-funded fast-follower can’t clone by Tuesday.</p><h3>Where it breaks, and the honest bit</h3><p>Credit to the review for saying where the mechanism fails, because most “build in public” gospel doesn’t. Regulated and enterprise infrastructure: a hospital procurement officer does not care about your Twitter. At scale, structurally: multiple product lines and a sales org eventually need a brand that doesn’t live or die on one person’s posting cadence. And when it’s performed instead of lived: people detect a staged “day in the life” even when they can’t articulate why.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*pN754uJYRnokSfyb.png" /></figure><p>That first exception matters for me specifically, because half my portfolio sells into enterprises and clinics, where the buyer wants institutional trust, not my feed. So the honest synthesis is: the founder channel is how you get the <em>first</em> believers and the first ten customers, and it is not how you get the hospital. Both are true. Plan for both and don’t run into the same ‘build before sell’ hole as articulated in <a href="https://medium.com/@hardik.goel214/34b2393a0ac8">Part-1</a> of the series.</p><h3>What I’m actually going to do</h3><p>Since the review demands specificity, here’s mine. My scar isn’t “passionate about AI.” It’s: two decades building data systems, a year of shipping products faster than I could sell them, and the slow, expensive realization that the launch was never the constraint. That’s a checkable, slightly embarrassing, entirely mine sentence, and it passes the swap-the-logo test, because no competitor would <em>want</em> to paste it.</p><p>The <a href="https://medium.com/@hardik.goel214/34b2393a0ac8">Part 1</a> caveat comes due here: launching wide divides your voice by the number of tools. The Founder-Is-the-Channel model only works with one identity, repeated. My own rule says one identity for twelve months. I wrote the rule. Time to obey it.</p><p>Next, Part 3, on the trap that catches you <em>after</em> distribution starts working: mistaking loud signals for real ones. The handbook has a ladder. Bring your waitlist. We’ll see where it sits.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems, and one uncomfortable year learning that “outbound is oxygen” is easier to write on a wall than to breathe. This is Part 2 of 4 on IBR’s handbook. Part 3, The Signal Cost Principle, is the framework I wish I’d had before my first “we broke the site” screenshot. Subscribe; the scar tissue economy pays out slowly, but it does pay.</em></p><p><strong>Series map:</strong> <a href="https://medium.com/@hardik.goel214/34b2393a0ac8">Part 1 · The Inverted Launch</a> → Part 2 · The Founder Is the Channel (this) →<a href="https://medium.com/@hardik.goel214/82c90bd36a9b"> Part 3 · The Signal Cost Principle</a> → <a href="https://medium.com/@hardik.goel214/d78b011a34ac">Part 4 · The Contribution Factor.</a></p><p><strong>Source:</strong> Ikana Business Review, <em>Mini Strategy Handbook for AI Founders</em> (August 2026), Review 02, “The Founder Is the Channel” by Richa Sati, including the polish inversion, the Swap-the-Logo Test, and the SCAR test; the full argument is at ikanabusinessreview.com. The confessions and the deflation framing are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=28d0d012eac1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You Don’t Win by What You Ship. You Win by What You Cut.]]></title>
            <link>https://medium.com/@hardik.goel214/you-dont-win-by-what-you-ship-you-win-by-what-you-cut-34b2393a0ac8?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/34b2393a0ac8</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[founders]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-09-25T21:31:01.731Z</atom:updated>
            <content:encoded><![CDATA[<h4>Part 1 of 4 · Notes on Ikana Business Review’s Mini Strategy Handbook for AI Founders: launch → distribute → read the signal → decide.</h4><p><strong>This one’s about launching, and it’s a framework I have personally, embarrassingly, half-implemented.</strong></p><p>Ikana Business Review put out a Mini Strategy Handbook this month, four field-tested frameworks for AI founders, and I read it the way you read a doctor’s report about yourself: nodding, wincing, occasionally saying “well, yes, but.” It’s good. Good enough that I’m going to walk through it in four parts, in the order the handbook intends, launch, distribution, traction, growth, and, in my usual manner, agree with the substance while arguing with a word or two.</p><p><strong>Part one</strong> is their first review: <strong>The Inverted Launch for AI</strong>, or ILA, from Vivek Bisht. Here’s the idea in one breath. For twenty years the playbook was sequential: ship one focused feature, validate it, layer more on over years. That worked because building was expensive and pivoting was slow. AI collapsed the cost of building, so the winning move flipped: ship a <em>cohort</em> of tools at once, watch which ones users actually pull on, cut the rest without sentiment, and reinforce the survivors. The unit of launch stops being the feature. It becomes the portfolio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*glGy2a2JJYHtRkqu.gif" /></figure><p>I have a confession that makes me the ideal reader of this review. I’ve shipped, this year, something north of eight products. Trinetra, Gambit, Helm, Faceoff, Sprout, Serum, etc. a couple of other items as well. I have run the first half of ILA with what can only be called enthusiasm. So take what follows as coming from someone who has field-tested the “launch wide” instruction and is still working on the “converge inward” part.</p><h3>The Arc story, which is really a telemetry story</h3><p>The handbook’s proof case is The Browser Company, and it’s a good one. Arc had accumulated Spaces, Easels, Boosts, Little Arc, Air Traffic Control, Live Folders, Calendar Preview. Power users adored it. Most users got lost. And the telemetry, once anyone was brave enough to read it, was brutal: only around five and a half percent of daily users regularly used more than one Space, about four percent used Live Folders, and Calendar Preview on hover was used by 0.4%. Meanwhile, two capabilities, chat-with-tabs and personalization, were used by roughly 40% and 37% of daily users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*6fpSjBa8myelku8p.gif" /></figure><p>So they put Arc into maintenance, rebuilt around the two things people actually used, and called it Dia. Atlassian bought the result for $610 million within a year of the prune. The handbook’s line, and it deserves the frame: sometimes the cut isn’t three features, it’s the whole product surface.</p><p>Notice what the story is <em>not</em>. It isn’t “they built the wrong thing.” Arc’s features were beautiful and people who used them loved them. It’s “they kept the wrong things too long.” That’s the entire inversion in one sentence, and it’s the sentence I’d tattoo on every founder who’s currently defending a feature because it demoed well.</p><h3>Where I’d argue with a word: it’s not really about launching</h3><p>Here’s my quibble, and it’s a friendly one. The framework is called the Inverted <em>Launch</em>, but launching wide is the easy half. Any founder with Claude Code and a weekend can ship five tools now; I have the repositories to prove it. The launch got cheap for everyone, which means launching wide isn’t an edge, it’s the entry fee.</p><p>The hard part, the part that actually inverts, is the <em>roadmap</em>. The old roadmap was addition: ship more over time. The new one is subtraction: cut what doesn’t earn pull. And subtraction is psychologically expensive in a way addition never was, because every tool you kill has a builder who loved it, a slide it appeared on, and a sunk cost that whispers “give it one more sprint.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*V7AADfPhUlMGvu13.png" /></figure><p>The table’s last row is the one that matters. Old: you win by what you ship. New: you win by what you cut. Which means the scarce skill in AI product-building is no longer engineering velocity, everyone has that now, it’s the willingness to read two to four weeks of live data and let it outrank your conviction, your investor’s opinion, and your ego. Building got cheap. Deciding didn’t.</p><h3>Steal this: the prune, made concrete</h3><p>The handbook gives five steps, launch wide, instrument everything, read the signal, cut, reinforce, and they’re right. I’d add teeth to step four, because in my experience step four is where founders quietly stop. Before you keep any tool past its trial window, ask three questions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*0KpjWI1kPK2zQIWL.png" /></figure><p>Would you build it today, from zero, knowing what the data says? If no, it’s sunk cost with a UI. Does it move the pull metric, or just the roadmap slide? If the slide, cut. And who’s loudly defending it: users, or the person who built it? If the builder, cut, kindly, and buy them a coffee, because they’ll build the next winner.</p><p>One caveat the handbook doesn’t dwell on and I will, since <a href="https://medium.com/@hardik.goel214/28d0d012eac1">Part 2</a> depends on it: launching wide has a distribution cost. Five tools means five stories to tell, five audiences to find, and a founder attention span divided by five. The cohort launch works only if you’re ruthless enough to converge fast, because until you do, your marketing is as diluted as your roadmap. Which is a fair description of my own year, and precisely why <a href="https://medium.com/@hardik.goel214/28d0d012eac1">Part 2</a>, on distribution, is the one I’m most nervous to write.</p><p>You don’t win by what you ship. I’ve shipped plenty. Time to go win by what I cut.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems and, this year, testing the “launch wide” half of this framework more thoroughly than anyone should. Next up, Part 2: The Founder Is the Channel, on why distribution at seed stage is a person problem wearing a marketing costume, written by a person with that exact problem. Subscribe so you don’t miss me being uncomfortable in public.</em></p><p><strong>Series map:</strong> Part 1 · The Inverted Launch (this) → <a href="https://medium.com/@hardik.goel214/28d0d012eac1">Part 2 · The Founder Is the Channel </a>→ <a href="https://medium.com/@hardik.goel214/82c90bd36a9b">Part 3 · The Signal Cost Principle </a>→ <a href="https://medium.com/@hardik.goel214/d78b011a34ac">Part 4 · The Contribution Factor.</a></p><p><strong>Source:</strong> Ikana Business Review, <em>Mini Strategy Handbook for AI Founders</em> (August 2026 edition), Review 01, “The Inverted Launch for AI (ILA)” by Vivek Bisht. The Arc/Dia telemetry figures and the Atlassian acquisition are as reported there; the full breakdown is at ikanabusinessreview.com. Commentary, the three prune questions, and the distribution caveat are mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34b2393a0ac8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SpaceXAI Will Rent You Five Employees for $200.]]></title>
            <link>https://medium.com/@hardik.goel214/spacexai-will-rent-you-five-employees-for-200-6768e695a372?source=rss-1f6541a8c84a------2</link>
            <guid isPermaLink="false">https://medium.com/p/6768e695a372</guid>
            <category><![CDATA[elon-musk]]></category>
            <category><![CDATA[grok]]></category>
            <category><![CDATA[hiring]]></category>
            <category><![CDATA[spacex]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Hardik Goel]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 21:31:01 GMT</pubDate>
            <atom:updated>2026-09-25T21:31:01.707Z</atom:updated>
            <content:encoded><![CDATA[<h3>SpaceXAI Will Rent You Five Employees for $200. They Share One Laptop, One Login, and One Set of Keys. It Says So in the Manual.</h3><h4>Grok Bot is worth hiring, five times over. It’s also the cleanest case study I’ve seen of a product whose marketing and its own documentation describe two different security models.</h4><p>On August 11, SpaceXAI, the xAI that came out of the SpaceX merger, shipped Grok Bot: always-on “AI teammates” you message like colleagues. Each gets a name, signs into your real tools, works while you sleep, and only comes back when something needs your approval. Bundled into SuperGrok Heavy and Cursor Ultra at $200 a month. Five named workers who finish jobs inside your actual Zendesk, your actual CRM, your actual inbox, instead of handing you drafts to paste.</p><p>Then the company wrote the catch into its own documentation and, to its credit, left the page up.</p><p>The launch post says Bots “have their own computer.” The docs, published the same day, say all of your Bots share <em>one</em> persistent cloud computer, isolated to your account, not to any Bot, and, in a sentence that appears word-for-word on two separate pages, &lt;cite index=”36–1”&gt;”Do not use separate Bots as a security boundary.”&lt;/cite&gt;</p><p>That’s not a footnote. For a product whose entire pitch is <em>log in as you and act unsupervised</em>, the security model is the product. And here the vendor published two of them.</p><h3>Five hires, one laptop</h3><p>Picture the org chart the marketing implies: five employees, five laptops, five sets of logins, five people you could fire individually. Now picture the architecture the docs describe: five names sharing one machine, where files, browser sessions and command-line credentials sit in one place available across the whole roster.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*X4iPFrsGxGlKJpNa.gif" /></figure><p>The moment your Inbox bot signs into Zendesk, that session belongs to Research, Ops, Bookkeeper and Support too, because they’re all sitting at the same desk. And when you delete a bot? Soft delete. The docs are plain: because Bots share a computer, files and logins on that computer may remain. One operator who ran four bots on a single account for a week and counted what each could reach reported eleven signed-in apps, one browser profile, and that deleting a bot left all of it standing. The name leaves. The keys stay.</p><p>Notice what the word “teammate” did there. Personnel metaphors smuggle in security intuitions: an employee has their own laptop, their own credentials, their own offboarding. The metaphor promises isolation. The architecture doesn’t honour it. That’s the whole gap, and it’s the gap you have to manage yourself.</p><h3>The 103-day gap</h3><p>Here’s the part that made me put my coffee down. On April 30, the NSA, CISA and the cyber agencies of the UK, Canada, Australia and New Zealand jointly published their first guidance on agentic AI. Its central instruction:</p><p>&lt;cite index=”45-1”&gt;never grant agents broad or unrestricted access, especially to sensitive data or critical systems, and use agentic AI only for low-risk and non-sensitive tasks.&lt;/cite&gt;</p><p>It also warned, specifically, that scope creep cascades across agents: if Agent A fully trusts Agent B, compromising B affects A and everyone else.</p><p>One hundred and three days later, a product shipped whose default architecture is Agent A, B, C, D and E fully trusting each other, by construction, on one machine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*uB6dq67_embij-Rd.gif" /></figure><p>I’m not saying the product is reckless. I’m saying the timing is comic, in the way only enterprise software can be: the world’s six most serious cyber agencies write a memo, and the next thing to ship is its photographic negative, with the negative helpfully documented in the FAQ.</p><h3>Draw the blast radius before the second bot exists</h3><p>Since the docs tell you in writing that bots aren’t a boundary, you can do something most buyers never get to do: draw the blast radius <em>before</em> you’ve made it. Sign in for the one bot that needs the site. Then open the others and see what they reach. That session is theirs the moment it exists.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*fXBfdSUkkAv5DzfM.gif" /></figure><p>That’s the fundamental. Everything else is arithmetic. One sign-in equals the browser session plus every file on the machine plus whatever the terminal remembers, times every bot present and future. Bot #2 doesn’t join your team so much as arrive already inside your first bot’s perimeter.</p><p>And this is precisely why the product <em>pays</em>: one sign-in is why five names can finish work inside your real tools instead of emailing you drafts. The feature and the risk are the same fact wearing different hats. You don’t get to keep one and return the other.</p><h3>The launch page vs the docs, on one card</h3><p>The security model isn’t the only place the two surfaces diverge. There’s no per-bot spend cap yet, the audit view of what your bots did is “still coming,” and data governance, retention, training opt-out, deletion, runs on Cursor’s terms, not xAI’s, which matters when your vendor-risk questionnaire is addressed to the wrong company.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*EzebsDUVJOnmwnUb.png" /></figure><p>Credit where it’s due: the credential handling xAI <em>did</em> document is genuinely better than the category norm. Passwords, passkeys, 2FA, CAPTCHAs and payments hand control back to the human; a masked “secure secret request” exists that’s excluded from the transcript and never shown to the model. The docs even hedge their own model-based Auto Review with a sentence every vendor should be forced to write: it should complement, not replace, least privilege and explicit approval boundaries. Somebody on that docs team has been in the room when things went wrong.</p><p>But note where those good controls sit: on top of the shared machine. Approvals gate what a bot <em>does</em>. They don’t partition what the roster <em>has</em>. And an approval controls the proposed action; whatever already ran stays wherever it landed.</p><h3>The one-machine playbook</h3><p>So how do you actually hire five and sleep? Architect hat on, no sirens. Treat the roster as one identity, and design accordingly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*UBMwqQD-LVJSlrBc.png" /></figure><p>Give each bot its own account on each app, not a shared one, since the docs tell you in writing that bot separation isn’t a boundary. Put the stop line in every bot’s description, in plain English: “never send, pay, delete or reply externally; hand back to me.” Cap the spend outside the product, at the card or the workspace, because there is no bot-level cap yet. Keep the money and the customer replies in your own hands, and let the other four start from scratch each morning on work that cannot bite: triage, research, drafts, reconciliation, CRM updates from transcripts, the reversible stuff the launch post itself uses as examples. And when you retire a bot, revoke at the source, because deleting the name does not delete the session.</p><p>Five role descriptions I’d actually paste, all sharing one closing line: <strong>Inbox</strong> — “read, label, draft replies; never send.” <strong>Research</strong> — “browse, summarize, cite; write only to /research; never log into anything new.” <strong>Ops</strong> — “update CRM from transcripts; flag anything touching billing.” <strong>Bookkeeper</strong> — “reconcile and categorize; propose, never pay.” <strong>Support</strong> — “draft answers to the queue; you approve every send.” Closing line for all five: <em>anything not on this list, stop and hand it back.</em> Bots can message each other and share context in threads, so the handoff between them is a message, not a login, which is exactly how you want it.</p><h3>The uncomfortable take</h3><p>Here’s what I actually think, and it’s the version nobody’s tweeting. Everyone is asking whether the bots are good. They’re good; the demos are real, and five of them for $200 is absurd value on paper. But your real limit on Grok Bot was never model quality. It’s how many logins you’re willing to put on one machine that five semi-supervised workers share. That number is smaller than five for most businesses, and it should be.</p><p>The hiring was always the easy half. The hard half is the same as it’s been for every wave I’ve written about this month: the capability is cheap and gorgeous, and the scarce thing is the trust architecture around it. Grok Bot didn’t build that architecture. To its credit, it told you so.</p><p>Read the docs, not the launch page. Then draw the circle before the second hire.</p><p>-Hardik</p><p><em>I’ve spent about two decades building data and AI systems, and long enough in enterprise procurement to know that “their own computer” and “one computer per account” are not synonyms, no matter how friendly the mascot. I write about AI’s mechanics, its hype cycles, and the sentences vendors publish and hope you skip. Subscribe. Then go check what your bots can reach.</em></p><p><strong>Sources / further reading:</strong> xAI’s Grok Bot launch post (x.ai/news, Aug 11, 2026) and the Grok Bot documentation, in particular the overview, FAQ, and approvals-security-and-privacy pages at docs.x.ai/grok-bot, the primary source for the shared-computer architecture, the twice-stated boundary warning, soft-delete behaviour, and credential handling. Digital Applied’s launch-day matrix of marketing-vs-docs contradictions (Aug 11), the most careful side-by-side I found. “Careful Adoption of Agentic AI Services,” CISA, NSA, ASD’s ACSC, CCCS, NCSC-NZ, NCSC-UK (Apr 30, 2026; media.defense.gov), plus Mayer Brown’s and Cloud Security Alliance’s summaries. TechTimes, eesel and Reworked on pricing, tiers and the Cursor/Anysphere governance layer. The four-bots-for-a-week field notes are from an operator’s public account, not my own test; my advice is mine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6768e695a372" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>