<?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 Robert Dunie on Medium]]></title>
        <description><![CDATA[Stories by Robert Dunie on Medium]]></description>
        <link>https://medium.com/@robdunie?source=rss-4100839dd1f6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*0YOXqz_ebM7sW6fL5kv8Pw.jpeg</url>
            <title>Stories by Robert Dunie on Medium</title>
            <link>https://medium.com/@robdunie?source=rss-4100839dd1f6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 02:35:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@robdunie/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[Building Governance for AI Agents Exposes Fractures in How We Govern Today]]></title>
            <link>https://pub.towardsai.net/building-governance-for-ai-agents-exposes-fractures-in-how-we-govern-today-d36eb8fc7413?source=rss-4100839dd1f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/d36eb8fc7413</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[autonomous-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-safety]]></category>
            <dc:creator><![CDATA[Robert Dunie]]></dc:creator>
            <pubDate>Fri, 27 Mar 2026 04:36:02 GMT</pubDate>
            <atom:updated>2026-03-28T15:43:24.371Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>How separation of concerns, automated controls, and forward-looking assessment solve the same coordination problems for agents and humans</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uRC3JTrEJBvoYyVfRtVsCw.png" /></figure><p><em>This is a follow-up to </em><a href="https://medium.com/@robdunie/governing-the-ant-farm-a-governance-first-framework-for-multi-agent-software-delivery-29245fc14bd9"><em>Governing the Ant Farm: A Governance-First Framework for Multi-Agent Software Delivery</em></a><em>, published March 19, 2026. It assumes familiarity with the DOODA-loop lineage, leadership triad, progressive autonomy model, compliance floor, and findings loop described in that piece. The framework has evolved significantly since publication — this article is about the patterns that emerged in that evolution. Visit the </em><a href="https://github.com/rduni/venutian-antfarm"><em>https://github.com/rduni/venutian-antfarm</em></a><em> to explore the latest changes and get started.</em></p><p>When I introduced Venutian Antfarm, I described it as a governance-first framework for multi-agent software delivery. The ant farm metaphor was useful shorthand: individually simple agents producing complex collective behavior, but only when the colony has structure. What I didn’t fully appreciate at the time was how literally the metaphor would hold up.</p><blockquote>The ant farm isn’t just a cute metaphor — it’s a digital twin of a delivery organization.</blockquote><p>Not because we’re anthropomorphizing agents. Because we’re solving the same coordination problems that led real organizations to develop executive oversight, automated compliance, and capability assessment in the first place. The agents don’t need a CRO because they aspire to be human. They need a CRO because governing a fleet of autonomous workers producing artifacts in a compliance-sensitive domain presents the same structural challenges whether those workers are carbon-based or silicon-based.</p><p>The recent framework evolution crystallized governance patterns that make this concrete:</p><ol><li>independent oversight,</li><li>automated controls, and</li><li>forward-looking assessment.</li></ol><h3>Independent Oversight: Separate Governance Agents Manage the Floor</h3><p>Every enterprise architect has had the conversation about separation of concerns. In code, it’s the single responsibility principle. In organizational design, it’s why boards don’t write code and engineering teams don’t set compensation policy. The people who set the rules shouldn’t be the same people executing the work — not because of distrust, but because governance and execution require fundamentally different orientations.</p><p>Venutian Antfarm’s governance tier applies this principle directly. Seven Cx agents — CRO, CISO, CEO, CTO, CFO, COO, and CKO — set policy independently of the operational chain. They don’t build features. They don’t write tests. They don’t deploy. They define and enforce the constraints within which the operational tier delivers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dauSjEbKEllooqmkul1E4g.png" /></figure><p>This isn’t organizational cosplay. Each role maps to a real governance function that emerged in the production system where the framework originated: a PII-sensitive case management application running on self-hosted Kubernetes with a zero-knowledge data architecture. In that system, the CRO’s compliance floor includes rules like “no PII in the operational database” and “secrets managed through vault, never hardcoded.” The CISO performs threat assessments against a real attack surface. The CTO evaluates whether the tech stack choices agents are making align with the project’s architectural direction. These aren’t decorative — they’re structural requirements that exist because the domain demands them.</p><p>The digital twin parallel sharpens when you look at how these governance agents interact with delivery. Like enterprise security councils and architecture review boards, the Cx tier operates through structured interfaces — floor rules (non-negotiable MUSTs), targets (aspirational SHOULDs), and guidance (Tier 3 recommendations) — rather than inserting itself into daily execution. The agents that build and review code interact with these policy artifacts, not with the governance agents directly.</p><p>The CEO agent makes the digital twin concept most explicit — it’s described in the framework as “the digital twin of the implementer.” It maintains an independent trust-based pace assessment, produces executive briefs, and grants scoped autonomy — a chief executive who delegates operational decisions to a leadership team while retaining strategic oversight.</p><h4>How Floor Changes Work</h4><p>The multi-floor governance model is where the separation of concerns really proves its value. Any Cx officer can define a governance floor — a set of non-negotiable rules with deterministic enforcement. The compliance floor (guarded by the CRO) and the behavioral floor (guarded by the COO) are the initial instances, but the pattern supports any number of floors with identical machinery.</p><p>Floor changes go through a structured consultation process that mirrors how enterprise governance boards operate. When a change is proposed, the floor’s guardian dispatches the CRO to facilitate multi-round Cx consultation — conducted as a subagent so the deliberation doesn’t pollute the working context. Each Cx agent advocates for their domain impact. The CRO synthesizes positions, identifies conflicts, and facilitates additional rounds until positions stabilize. Only the compact result — risk assessment, Cx positions, recommendation — enters the main conversation context. The delivery team gets the decision; they don’t sit through the debate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xvkpF-tWVkQEuoD-O247bA.png" /></figure><h3>Automated Controls: Policy as Code for Agent Governance</h3><p>Separation of concerns is necessary but not sufficient. A governance body that sets policy but can’t enforce it is a suggestion box. In enterprise governance, this is the perennial gap between writing a SOX control and actually verifying that the control operates effectively — a gap that traditional organizations bridge with periodic manual audits, which means controls are verified quarterly at best and violated between audits at worst.</p><p>Venutian Antfarm closes this gap by applying the policy-as-code pattern to agent governance. The core idea: governance rules are authored as human-readable Markdown, then compiled into machine-enforceable artifacts. The same document a governance reviewer reads is the source of truth for automated enforcement. No translation layer. No synchronization problem between the policy binder and what’s actually running.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8oQJvhEE1iDIP7Wy7lkxNQ.png" /></figure><p>This is the digital twin at its most operational — not just mirroring organizational structure but automating the governance machinery that the structure is supposed to provide.</p><h4>What the Compiler Produces</h4><p>A floor compiler transforms governance Markdown into several categories of enforceable artifacts:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FvumTAmV-VSlwTU1y3ds0Q.png" /></figure><p>Both compilers are idempotent, dry-run capable, and share a utility library for manifests, verification, and structured logging.</p><h4>Why This Matters Beyond the Mechanics</h4><p>Here’s a concrete example from the production system. One compliance floor rule states that PII must never appear in application logs. When an agent writes code that feeds a client’s name into a log statement, a compiled hook blocks the commit before the LLM even sees the result. The agent can’t argue, rationalize, or hallucinate its way past the control. This is what distinguishes policy-as-code from policy-as-prompt: prompts are suggestions that agents interpret; compiled controls are deterministic boundaries that agents cannot cross.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/708/1*7amk1hRAowq1ZJlvoQX2GA.png" /></figure><p>The distinction matters enormously for trust. Progressive autonomy — letting agents earn expanded scope through demonstrated performance — only works if the floor beneath that autonomy is non-negotiable. You can give an agent Fly-pace autonomy precisely because you know the enforcement hooks will catch a PII leak regardless of how much latitude the agent has on everything else. The floor makes the ceiling safe.</p><p>A SOX auditor examining this system wouldn’t need to sample — every enforcement action is logged, every policy change is tracked through the consultation process, and every artifact’s integrity is cryptographically verifiable. Most enterprise governance programs would love to make the same claim about their human-operated controls.</p><h3>Adapting to What’s Coming: Forward-Looking Assessment of Complexity and Novelty</h3><p>Independent oversight sets the rules. Automated controls enforce them. But governance that only looks backward — “did you follow the rules?” — misses the harder question: “are you ready for what’s next?”</p><p>This is where the Cynefin-informed progressive autonomy model becomes more than a trust-building exercise. The Crawl–Walk–Run–Fly progression, as described in the original article, adjusts fleet autonomy based on evidence. What I want to emphasize here is that pace ratchets <em>down</em> when complexity surfaces, not just up when metrics look good. This is the dynamic OODA influence: the system continuously reorients based on observed terrain, and sometimes reorientation means slowing down.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XR5BXIWUqtMeAEQPBp9lTQ.png" /></figure><p>In the production case management system, we experienced this during a migration to a new authentication provider. The fleet was operating at Run pace — expanded autonomy, batch oversight, only strategic decisions requiring proposals. The auth migration introduced integration complexity that crossed domain boundaries: frontend session management, backend token validation, infrastructure certificate rotation, and compliance implications for audit logging. The triad consulted, and consensus was that the complexity warranted Crawl-level scrutiny for auth-related work items even though the fleet-wide pace remained at Run. Complexity overrides pace.</p><p>This is the organizational maturity model made dynamic. In most enterprises, capability assessments are periodic — annual reviews, quarterly architecture evaluations, maturity model benchmarks. The assessment happens, a score is assigned, and the organization operates at that assumed capability level until the next review cycle. The Venutian Antfarm equivalent is continuous: every significant problem triggers a fresh evaluation of whether the current governance posture matches the observed complexity.</p><h3>The Sensing Apparatus</h3><p>Three mechanisms feed this continuous evaluation:</p><p><strong>Triad consultation </strong>on significant problems surfaces multiple perspectives before pace decisions. The Product Owner brings delivery context, the Solution Architect brings technical assessment, the Scrum Master brings process evidence. High consensus and appropriate pace — proceed. Low consensus — each perspective surfaces independently for the human operator to decide. This mirrors how effective engineering leadership teams work: disagreement is surfaced, not suppressed.</p><p><strong>Cx assessments </strong>provide domain-specific forward looks. The CISO evaluates whether emerging threats require governance posture changes. The CTO evaluates whether technology direction increases or decreases the fleet’s ability to deliver safely. The COO monitors agent performance and triggers retraining when patterns suggest capability gaps rather than one-off errors. These aren’t backward-looking audits — they’re forward-looking assessments of readiness.</p><p>The <strong>rewards system </strong>creates behavioral signal. Governance and leadership agents issue kudos and reprimands scoped to their domain — a CISO reprimand on a shallow security review carries different weight than a PO kudo on delivery speed. Over time, agent behavioral profiles emerge from the pattern of signals across work items. Tension detection — when conflicting signals appear on the same work — surfaces disagreements that might indicate the work is at the edge of the fleet’s current capability. These tensions don’t resolve automatically; they surface for human judgment. The system is saying: “we’re not sure about this one, and the fact that we’re not sure is itself information.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6U-WOrRVreuXANTMFtExCQ.png" /></figure><p>This is what distinguishes the rewards system from a simple scorecard. It’s not measuring compliance — the enforcement hooks handle that deterministically. It’s measuring <em>judgment</em>. Are the agents making good calls in the spaces between the rules? Are they developing the kind of contextual reasoning that warrants expanded autonomy?</p><h4>The Cynefin Connection</h4><p>The Crawl–Walk–Run–Fly model maps loosely to Cynefin dispositions, as I described in the original article. What I want to add here is the question of <em>disorder</em> — the Cynefin domain where you don’t know which domain you’re in.</p><p>I think the framework equivalent is when the consultation process can’t reach resolution. When the CISO’s threat assessment contradicts the CTO’s technology direction, and neither position is wrong — they’re operating from genuinely different risk frames — the current design escalates: positions stabilize or they surface to the human as unresolved tensions. But at scale, the consultation process itself might need governance. What does a consultation circuit breaker look like? Is there a structural decomposition that keeps Cx deliberation tractable as floor count grows? The framework’s architecture is designed to support that exploration without rearchitecting the enforcement layer underneath.</p><h4>Does the Digital Twin Hold?</h4><p>The three patterns reinforce the digital twin framing from different angles. Oversight mirrors organizational structure. Controls mirror governance machinery. Assessment mirrors capability evaluation.</p><p>These patterns are grounded in practice — a 16-agent fleet delivering a real compliance-sensitive codebase with PII handling rules, a zero-knowledge data architecture, and a self-hosted Kubernetes deployment target. The constraints are real and the consequences for violation are real. Scaling beyond one fleet, one domain, and one operator will surface new challenges — but the digital twin framing gives us a useful lens for anticipating them. Scaling an autonomous agent fleet will likely hit the same coordination boundaries that scaling a human delivery organization does: governance overhead that grows faster than headcount, consultation processes that slow down as stakeholder count increases, and behavioral signals that lose resolution as the fleet diversifies. These are well-studied problems in organizational design. The patterns may need to evolve, but the structural parallels suggest we’re not starting from scratch.</p><p>Areas worth exploring:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FAgEdJhoTWzNWFtDBIBdCA.png" /></figure><h3>Looking Forward</h3><p>These patterns aren’t specific to agent fleets. Any organization adopting AI-augmented delivery will face the same coordination problems: who sets the rules, how are they enforced, and how do you assess readiness for increased autonomy?</p><p>The patterns that work for governing agents also work for governing the humans working alongside them — not because the tooling transfers, but because the underlying principles do: separate policy-setting from execution, treat policy as code rather than documentation, and evaluate capability trajectories rather than just current compliance.</p><p>The digital twin framing suggests that the relationship is bidirectional. Organizational patterns inform agent governance, but the precision required to make those patterns work for agents — the explicit consultation protocols, the deterministic enforcement, the structured behavioral feedback — reveals how imprecise the same patterns often are in human organizations. When you’re forced to codify “how does the security council interact with the delivery team,” you discover how much of that interaction was previously informal, inconsistent, and dependent on individual relationships.</p><p>Building governance for an agent fleet doesn’t just solve the agent coordination problem. It forces you to articulate — precisely, explicitly, enforceably — governance structures that most organizations leave informal. And once you’ve articulated them, you can’t unsee the gaps in the human version.</p><p><a href="https://github.com/rdunie/venutian-antfarm"><em>Venutian Antfarm</em></a><em> is open source. The </em><a href="https://github.com/rdunie/venutian-antfarm/releases"><em>release notes</em></a><em> cover the multi-floor governance model, compiler evolution, and rewards system in implementation detail. If you’re building governance patterns for agent fleets — or finding that the patterns from agent governance clarify your thinking about human organizational design — I’d welcome the conversation.</em></p><p><em>Rob Dunie is Principal and Owner of </em><a href="https://robdunie.com/"><em>RD Digital Consulting Services, LLC</em></a><em>, with over 25 years of experience spanning business strategy, enterprise architecture, and technology research. His previous work at Gartner on digital business moments and adaptive strategy informs his current focus on AI governance frameworks and multi-agent delivery systems.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d36eb8fc7413" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/building-governance-for-ai-agents-exposes-fractures-in-how-we-govern-today-d36eb8fc7413">Building Governance for AI Agents Exposes Fractures in How We Govern Today</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Governing the Ant Farm: A Governance-First Framework for Multi-Agent Software Delivery]]></title>
            <link>https://pub.towardsai.net/governing-the-ant-farm-a-governance-first-framework-for-multi-agent-software-delivery-29245fc14bd9?source=rss-4100839dd1f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/29245fc14bd9</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agent-harness]]></category>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Robert Dunie]]></dc:creator>
            <pubDate>Thu, 19 Mar 2026 06:30:11 GMT</pubDate>
            <atom:updated>2026-03-21T18:10:30.158Z</atom:updated>
            <cc:license>https://creativecommons.org/licenses/by-nd/4.0/</cc:license>
            <content:encoded><![CDATA[<h4><em>How the Dynamic OODA loop, release train engineering, and structured learning loops converge in a practical harness for AI agent fleets</em></h4><blockquote>This article describes the initial version of the Venutian Antfarm framework. The framework is under active development and constantly being enhanced with new capabilities, agents, and operational tooling. Visit the <a href="https://github.com/rduni/venutian-antfarm">https://github.com/rduni/venutian-antfarm</a> to explore the latest <br> changes and get started.</blockquote><h3>The promise of agentic AI in software delivery is seductive: autonomous agents writing code, running tests, deploying to production. The reality, for anyone who has actually tried to orchestrate multiple agents on a non-trivial codebase, is considerably messier. Agents hallucinate. They drift from architectural intent. They make locally rational decisions that are globally incoherent. Without governance, a fleet of AI agents isn’t a team — it’s an ant farm with no colony structure.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6kRZs_KUedABgyGZ_E_hzw.png" /></figure><p>This is the problem that led me to build <a href="http://github.com/rdunie/venutian-antfarm">Venutian Antfarm</a>, an open-source framework for structured multi-agent software delivery with progressive autonomy, evidence-based governance, and measurable quality control. But before I walk through what the framework does and how it works, I want to share the thinking that produced it — because the architecture is inseparable from the governance philosophy underneath.</p><h3>From OODA to DOODA: Sensemaking as the Governance Bottleneck</h3><p>My thinking on agent governance traces back to research I co-authored at Gartner on digital business moments — published as “Get Out of Strategic Limbo: How to Discover and Exploit the Business Moments That Drive Digital Transformation.” ¹ That work used the Dynamic OODA (DOODA) loop — adapted from Brendt Brehmer’s amalgamation of Boyd’s OODA loop with cybernetic command-and-control theory² — as the core model for connecting business strategy to real-time detection and fulfillment of transient opportunities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Im4Zh-Jne1IBMkaMsJVhqA.png" /><figcaption>The DOODA loop combines stategic sense-making with iterative decision making and action.</figcaption></figure><p>Boyd’s original OODA loop (Observe-Orient-Decide-Act) was a model of competitive decision-making where Orient is the <em>schwerpunkt</em> — the center of gravity.³ Brehmer’s contribution was to make the loop dynamic: rather than discrete cycles, the phases operate continuously and concurrently, with feedback from effects and friction informing the next sensemaking pass even while actions from the previous cycle are still in flight.</p><p>In our Gartner research, we mapped business moments along two axes — whether triggers are known or unknown, and whether paths to outcome are known or unknown — producing four dispositions drawn from the Cynefin framework:⁴ standard operating procedures (simple), opportunistic moments (complicated), improvisational moments (complex), and innovative moments (chaotic). Each disposition demands a different sensemaking tactic. Crucially, a moment’s disposition can shift mid-execution — what starts as opportunistic can become innovative when assumptions fail.</p><blockquote>An agent fleet’s sensemaking function — its ability to orient on the current state of the codebase, the delivery process, and the business intent — is the critical bottleneck.</blockquote><p>Individual agents can act with impressive speed, but without continuous reorientation at the fleet level, those actions drift from coherence. The DOODA loop taught me that governance isn’t a gate you pass through; it’s a persistent orientation function that runs concurrently with execution.</p><p>This concept connected directly to my experience implementing SAFe Agile Release Trains. In SAFe, the ART leadership triad — Product Management, Enterprise Architect, and Release Train Engineer — collectively maintains orientation across multiple scrum teams working in parallel on shared codebases.⁵ None of them writes code. Together, they <em>are</em> the sensemaking function — and they serve as a useful, if imperfect, analog for the multi-agent coordination problem we now face with AI.</p><blockquote>What does the agile governance look like when the teams are AI agents?</blockquote><h3>The Triad: Servant Leadership as Governance Architecture</h3><p>The answer begins with a structural observation. In well-functioning agile delivery, governance isn’t imposed through command-and-control hierarchies — it emerges from the interplay of three distinct perspectives: business value, technical soundness, and process health. In SAFe, these map to the Product Manager, Enterprise Architect, and RTE roles. In Scrum, it’s the Product Owner, the development team’s technical leadership, and the Scrum Master.⁵</p><p>Venutian Antfarm formalizes this as the Leadership Triad: a Product Owner agent, a Solution Architect agent, and a Scrum Master agent. These three operate as servant leaders — they don’t dictate what execution agents do, they enrich the context in which execution agents make decisions. The PO validates business value and acceptance criteria. The SA ensures technical coherence, manages non-functional requirements, and identifies cross-system dependencies. The SM owns process health, pace control, and conflict mediation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/745/1*P4_7SrscAc8_tUWgcOlCcg.png" /></figure><blockquote><em>The triad doesn’t command — it enriches. When specialists ask frequent questions, that’s a signal to enrich the context, not tighten control.</em></blockquote><p>The triad operates through consensus protocols. For conservative actions (maintaining or tightening governance), unified recommendation is sufficient. For accelerations (expanding autonomy, increasing pace), triad unanimity is required before seeking operator approval. When the triad disagrees, all perspectives surface for human decision-making. This isn’t bureaucracy — it’s checks and balances. The same structural principle that makes democratic governance resilient makes agent fleet governance resilient.</p><p>Below the triad sits the full fleet structure — harness agents providing governance, app-layer specialists doing the work, and a human operator providing evidence-based oversight:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/803/1*AzVk9FtfcvqBZnPnx9H-Mw.png" /><figcaption>The four-layer fleet architecture. The harness layer ships with the framework; the app layer is yours to define. The Human Operator engages through evidence-based oversight, not continuous monitoring.</figcaption></figure><h3>Progressive Autonomy: Trust as a Measurable Gradient</h3><p>Most agent frameworks offer a binary choice: approve every action, or let agents run free. This mirrors a false dichotomy that plagues human organizations too — micromanagement versus abdication. Neither works at scale.</p><p>Venutian Antfarm replaces this binary with a four-stage autonomy progression: Crawl–Walk–Run–Fly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FlWQ47bQIhgPnUCVLv9W_A.png" /></figure><p>Crawl is the starting state. Agents propose nearly everything and explain their reasoning. The human reviews and approves. This is expensive, slow, and absolutely necessary — it’s how the system and the human build mutual calibration. Think of it as sprint zero for your agent fleet.</p><p>Walk introduces standard three-tier autonomy: some actions are autonomous (reading code, running tests), some require proposals (cross-domain changes, new dependencies), and some require escalation (compliance implications, destructive operations).</p><blockquote>The human engages at milestone reviews rather than every action.</blockquote><p>Run expands autonomy further. Agents chain work across items, batch oversight replaces continuous review, and only strategic calls require proposals. This is where throughput starts to approach the theoretical promise of multi-agent delivery.</p><p>Fly grants full autonomous operation. Agents execute, commit, and deploy with metrics-driven monitoring rather than action-level approval. But critically, even at Fly pace, the escalation tier never disappears. Compliance floor violations, strategic decisions, and destructive operations always surface to a human.</p><blockquote>Transitions are evidence-driven and reversible. You don’t set pace based on vibes.</blockquote><p>In our implementation, run pace requires a change failure rate below 5%, first-pass handoff acceptance above 95%, and declining findings frequency — thresholds that teams should calibrate to their own domain and risk tolerance. Pace ratchets back down when problems emerge. When significant problems surface at any pace, the triad consults: if high consensus and pace is current, proceed; if high consensus and pace should increase, a unified recommendation goes to the user; if low consensus, each perspective surfaces independently for the human operator to decide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uwL62jykkYpe4vVK0rvMxA.png" /><figcaption><em>The full Crawl–Walk–Run–Fly model with evidence gates and triad consultation. Significant problems at any pace trigger triad review rather than automatic regression.</em></figcaption></figure><p>This is where the DOODA influence shows most clearly. Crawl corresponds to the chaotic/innovative Cynefin disposition where you must act-sense-respond because nothing is yet proven; Fly corresponds to the simple/standard-operating-procedure disposition where triggers and paths are well understood. Cynefin practitioners will recognize the critical danger here: the cliff between simple and chaotic.⁴ When a system operates in the simple domain too long, it risks complacency — oversimplifying situations, failing to account for exceptions, and suddenly tipping into chaos when an unrecognized edge case breaks assumptions. The escalation architecture is the structural mitigation.</p><h3>The Findings Loop: Structured Learning With a Fitness Function</h3><p>Rather than passive memory accumulation — the approach most agent frameworks take, where you dump everything into a growing context — Venutian Antfarm implements a structured continuous improvement loop with a measurable fitness function, borrowing the shape of reinforcement learning without the algorithms.⁶</p><p>The cycle works as follows:</p><ol><li>Signal generation. Agents record notable events during work — surprises, boundary tensions, pattern observations, replicable successes, outright failures — as findings with urgency classification (critical through low).</li><li>Curation. The Scrum Master agent reviews findings, identifies patterns, and proposes refinements: prompt updates, memory enrichments, autonomy tier adjustments, or process changes.</li><li>Application. The triad reviews and applies approved refinements across the fleet.</li><li>Measurement. The same finding type should decrease in frequency over time. If it doesn’t, the refinement failed and needs revision.</li></ol><p>The structural shape resembles reinforcement learning — findings act as the signal, curation evaluates policy, refinement updates it, measurement closes the loop — but this is deliberate analogy, not implementation. There are no gradient updates or reward models. What makes it effective is the human-in-the-loop component at steps 2 and 3, preventing the objective drift that purely automated learning systems can produce.</p><p>The findings loop also provides the evidential foundation for pace transitions. You can’t credibly argue for Run pace if the findings register shows recurring architectural drift or compliance near-misses. The data makes the case, not the aspiration.</p><h3>The Compliance Floor: Governance That Doesn’t Negotiate</h3><p>Progressive autonomy creates a risk: what happens when a high-pace fleet encounters a situation that should never be handled autonomously? The answer is the compliance floor — a small set (typically 3–5) of non-negotiable rules that override all autonomy tiers, pace settings, and priority decisions, enforced by a dedicated Compliance Auditor agent during review phases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/657/1*fKAt82Gwcq0U9sFHcCfasA.png" /></figure><blockquote><em>Pace Control, Autonomy Tiers, and Priority Decisions are all overridden by the Compliance Floor. The rules are yours to define; the override behavior is non-negotiable.</em></blockquote><p>In the production system where this pattern originated — a PII-sensitive case management application — the compliance floor included rules like “no PII in the operational database,” “row-level access control on every query,” and “secrets managed through vault, never hardcoded.” An agent at Fly pace still escalates on compliance floor contact. The floor is the governance bedrock that makes progressive autonomy safe.</p><h3>Measuring What Matters: DORA Meets Flow Quality</h3><p>You can’t govern what you can’t measure. Venutian Antfarm implements a metrics pipeline built around two complementary frameworks.</p><p>DORA metrics capture delivery outcomes: deployment frequency, lead time from promotion to acceptance, change failure rate, and mean time to recovery. These metrics were empirically validated as the strongest predictors of organizational software delivery performance,⁷ and they translate directly to agent fleet measurement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/463/1*dWgT8rcBtDu7QRfS1sCxyA.png" /></figure><p>Flow quality metrics capture process health: first-pass yield (percentage of handoffs accepted without rejection), rework cycles, task abandonment rates, blocked time, and restart rates. First-pass yield by agent boundary pair is particularly actionable — it reveals exactly where inter-agent handoff quality breaks down.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*--FZg-67Bx0Hi0ui7C8zGw.png" /><figcaption><em>The full metrics pipeline. Agent events are logged append-only to events.jsonl. The dashboard splits them into DORA outcomes and flow quality signals. The SM produces pace change recommendations; the PO adjusts priorities.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gtPLgpegMnUC0ujeRqFmQA.png" /><figcaption><em>Simplified view: ops/metrics-log.sh logs events; ops/dora.sh computes DORA and ops/dora.sh –flow computes flow quality; the SM issues a pace recommendation.</em></figcaption></figure><p>Together, these metrics feed back into both pace control decisions and findings curation, creating the continuous orientation loop that the DOODA model prescribes.</p><h3>The Architecture: Six Agents You Don’t Write</h3><p>The framework ships with six governance-layer agents: the Product Owner, Solution Architect, and Scrum Master (the triad), plus a Memory Manager (knowledge consistency and staleness detection), Platform Ops (DORA metrics, CI/CD visibility, deployment health), and Compliance Auditor (domain-specific rule enforcement).</p><p>Users extend these with custom specialist agents tailored to their technology stack and domain. The framework uses an inheritance model where app-layer agents override harness agents only in specified fields — unmodified governance properties are preserved. This means you get structured governance without duplicating governance logic into every specialist agent you write.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5h9z8iEARMDULj6-xRpH8g.png" /><figcaption><em>Agent inheritance: the harness agent provides base properties (name, model, protocol, autonomy); the app agent extends with overrides (retro cadence, autonomy override); the runtime agent is the merged result. Governance properties not explicitly overridden are preserved.</em></figcaption></figure><p>The framework uses a three-layer enforcement architecture, a design choice motivated by the observation that not every governance check needs an LLM invocation:</p><ul><li>Hooks are deterministic shell scripts triggered on tool events. They block dangerous operations, auto-format code, and detect drift at zero LLM cost.</li><li>Memory provides persisted behavioral rules that carry across sessions at minimal token cost on recall.</li><li>Skills and checklists are loaded prompts with mandatory steps for complex workflows — higher token cost but high reliability through forced sequencing.</li></ul><p>Agent dispatching also includes an isolation model: agents can be dispatched against a temporary worktree (assessing committed state only) or the live working tree (current state), controlled by a frontmatter flag.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/550/1*lAqOK3Tphr7oqKNqRlNmDA.png" /><figcaption><em>Worktree isolation prevents review-phase agents from accidentally mutating working state. The flag in frontmatter determines which view the agent reads.</em></figcaption></figure><p>Within each task, agents select models appropriate to the work: synthesis and judgment tasks use expensive models with medium thinking budget; structured work with clear patterns uses mid-tier; mechanical checks use cheap models with no thinking overhead. When an agent hits its thinking ceiling, that’s a findings signal — the task was mis-classified or context needs enriching.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/814/1*ezThFHE-XcGkzGcus9LNiA.png" /><figcaption><em>Model selection decision tree. Cost-appropriate model invocation is itself a governance artifact — thinking ceiling hits are diagnostic signals fed back into the findings loop.</em></figcaption></figure><p>Agent Cost Analysis tracks token spend across every agent in the fleet — broken down by agent, model tier, and work item. It answers the question every team running agentic workflows eventually asks: where are my tokens going, and is the spend producing quality? The dashboard surfaces per-agent call counts, model distribution against budget ceilings, and per-item token cost so you can spot misclassified tasks, stale context driving re-exploration, and whether your cost-per-accepted-item trend is improving over time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/544/1*OWwB7IjZJdSwF4qywDgjKA.png" /></figure><p>Work items flow through a nine-phase lifecycle — Groom, Promote, Build, Review, Fix, Deploy, Accept, Retro, Checkpoint — with shift-left validation at every stage. The executing agent owns the full validation cycle, not just their piece. This eliminates the “throw it over the wall” anti-pattern that plagues both human teams and naive agent architectures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mWBb01d5ZT3__SY6klTlvA.png" /><figcaption><em>The nine-phase lifecycle in full detail. Note the rework loop from Review back to Fix, the pass path through Deploy to Accept, and the feedback arc from Retro into the next Groom.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n34c6jg0Q2hgLTOpJ52BpQ.png" /><figcaption><em>Simplified lifecycle overview. Checkpoint at phase 9 monitors process health across the full cycle and feeds into the next item’s Groom.</em></figcaption></figure><h3>Related Work</h3><p>The governance challenge for AI agent fleets is attracting attention from multiple directions, each addressing a different slice of the problem.</p><p>Dong et al.’s Governance-as-a-Service (GaaS) framework⁸ introduces runtime policy enforcement through trust scoring and graduated response modes — block, warn, or escalate — applied to individual agent outputs. Venutian Antfarm operates at a different level of abstraction: rather than filtering per-output, it governs the delivery process itself through fleet-wide pace control, role-based sensemaking, and lifecycle metrics. The two approaches are complementary — GaaS could enforce output-level policy within a fleet that Venutian Antfarm governs at the process level.</p><p>Snowden’s Cynefin framework⁴ has been applied to generative AI contexts with the argument that the domain is inherently complex and demands learning amplifiers rather than rigid policy factories. Our Crawl–Walk–Run–Fly progression and triad consultation model are an attempt to operationalize that insight — treating governance posture as a dynamic response to observed complexity rather than a fixed configuration.</p><p>The DORA research program⁷ provides the empirical foundation for our delivery outcome metrics. Wang et al.’s survey of LLM-based autonomous agents⁹ documents the structural evolution from individual ungoverned agents toward coordinated multi-agent assemblies — the trajectory that progressive autonomy with evidence-gated transitions is designed to support.</p><h3>Why “Venutian Antfarm”</h3><p>The name is deliberately playful. An ant farm is a system where individually simple agents produce complex collective behavior — but only when the colony has structure. The “Venutian” qualifier (yes, it’s a deliberate misspelling) nods to the alien, sometimes surreal experience of watching an AI agent fleet coordinate on real software — and to Alien Ant Farm, the band whose most famous track was a cover of someone else’s work, reinterpreted with enough energy to make it their own. That’s not a bad metaphor for what a governed agent fleet does: takes established patterns and executes them with a different kind of intensity.</p><p>The name is a reminder not to take ourselves too seriously while doing serious work. <em>Annie are you OK?</em></p><h3>Getting Started</h3><p>Venutian Antfarm is open source: <a href="https://github.com/rdunie/venutian-antfarm">github.com/rdunie/venutian-antfarm</a></p><p>Clone the repo, define your compliance floor, add specialist agents for your stack, configure your deployment targets, and open in Claude Code. The six governance agents activate immediately alongside your specialists. The framework is dual-licensed under AGPL 3.0 (with an app-layer exemption) and a commercial license — your agents, compliance rules, and configurations remain your intellectual property.</p><p><em>Rob Dunie is a product and technology innovation leader with experience spanning enterprise architecture, agile delivery, and agentic software development. His previous research at Gartner on digital business moments and adaptive strategy informs his current work on AI governance frameworks.</em></p><p><em>If you’re experimenting with multi-agent delivery systems, I’d love to hear what governance challenges you’re encountering. The problems are real, the solutions are evolving, and the conversation is just getting started.</em></p><h3>Appendix A: Framework Comparison</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/987/1*pz_tWugD10GcroEC_T9guQ.png" /></figure><h3>Appendix B: Research Gaps &amp; Future Work</h3><p><strong>Progressive autonomy thresholds</strong>. Our evidence gates use domain-inspired thresholds (CFR &lt; 5%, FPY &gt; 95%). How do optimal values vary by codebase complexity, domain risk, or LLM backend? Controlled benchmarks against ungoverned baselines could quantify the throughput-safety tradeoff empirically.</p><p><strong>Triad effectiveness</strong>. Does triad consensus reduce hallucination and architectural drift more than solo governance agents, or does the coordination overhead negate the benefit? A/B tests on SWE-bench variants measuring handoff yield and escalation frequency would clarify this. ⁶</p><p><strong>Findings loop impact</strong>. Does declining findings frequency correlate with sustained Fly-pace stability, or does it produce the Cynefin “cliff” — a fleet so operationally confident it becomes brittle to novel edge cases? ⁴ Longitudinal studies across multiple fleets are needed.</p><p><strong>Metrics generalization</strong>. DORA and flow quality metrics were developed for human teams. ⁷ Do agent-specific failure modes — deadlock, prompt drift, context diffusion — require new instruments not captured by existing frameworks?</p><p><strong>Scaling limits</strong>. At what fleet size does triad coordination overhead outweigh benefit? Governance decay in large swarms — where triad agents can no longer maintain coherent orientation across the full portfolio — remains an open question. ⁹</p><p>I welcome empirical contributions via the repo — your production data could shape v2.</p><h3>Notes</h3><ol><li><a href="https://ito.hkbu.edu.hk/services/gartner.html">Scheibenreif, D. and Dunie, R.</a> “Get Out of Strategic Limbo: How to Discover and Exploit the Business Moments That Drive Digital Transformation.” Gartner Research, November 2017. ID G00341410.<br><em>(Gartner paywalled; institutional gateway)</em></li><li><a href="http://www.dodccrp.org/events/10th_ICCRTS/CD/papers/365.pdf">Brehmer, B.</a> “The Dynamic OODA Loop: Amalgamating Boyd’s OODA Loop and the Cybernetic Approach to Command and Control.” Proceedings of the 10th International Command and Control Research and Technology Symposium, 2005.<br><a href="https://ooda.de/media/berndt_brehmer_-_the_dynamic_ooda_loop.pdf"><em>Alternate mirror</em></a></li><li><a href="https://media.defense.gov/2018/May/22/2001920668/-1/-1/0/B_0151_Boyd_Discourse_Winning_Losing.pdf">Boyd, J.</a> “The Essence of Winning and Losing.” Unpublished briefing slides, 1996.<br><em>(Closely related compiled material: A Discourse on Winning and Losing)</em></li><li><a href="https://hbr.org/2007/11/a-leaders-framework-for-decision-making">Snowden, D. and Boone, M.</a> “A Leader’s Framework for Decision Making.” <em>Harvard Business Review</em>, November 2007.</li><li><a href="https://scaledagileframework.com/agile-release-train/">Scaled Agile, Inc.</a> <em>SAFe 6.0 Framework: Agile Release Train.</em> scaledagileframework.com, 2023.<br><a href="https://scaledagile.com/certification/safe-release-train-engineer/"><em>RTE role</em></a></li><li><a href="https://arxiv.org/abs/2203.02155">Ouyang, L. et al.</a> “Training language models to follow instructions with human feedback.” <em>NeurIPS</em>, 2022.</li><li><a href="https://www.simonandschuster.com/books/Accelerate/Nicole-Forsgren-PhD/9781942788331">Forsgren, N., Humble, J., and Kim, G.</a> <em>Accelerate: The Science of Lean Software and DevOps.</em> IT Revolution Press, 2018.<br><a href="https://www.devops-research.com/"><em>DORA research</em></a></li><li><a href="https://arxiv.org/abs/2508.18765">Dong, Y. et al.</a> “Governance as a Service: Runtime Policy Enforcement for LLM Agents.” arXiv:2504.01234, 2025.</li><li><a href="https://arxiv.org/abs/2308.11432">Wang, L. et al.</a> “A Survey on Large Language Model based Autonomous Agents.” <em>Frontiers of Computer Science</em>, 2024.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29245fc14bd9" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/governing-the-ant-farm-a-governance-first-framework-for-multi-agent-software-delivery-29245fc14bd9">Governing the Ant Farm: A Governance-First Framework for Multi-Agent Software Delivery</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Update — March 2026: Winners, Losers, and the “Technology‑Neutral” Future of Federal R&D]]></title>
            <link>https://medium.com/@robdunie/update-march-2026-winners-losers-and-the-technology-neutral-future-of-federal-r-d-28a88db1bccf?source=rss-4100839dd1f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/28a88db1bccf</guid>
            <category><![CDATA[sttr]]></category>
            <category><![CDATA[sbir]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Robert Dunie]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 18:12:06 GMT</pubDate>
            <atom:updated>2026-03-16T18:12:06.408Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fVGiUp__LplxCAuXmKAa5A.jpeg" /></figure><p>Since my February 25 post, the SBIR and STTR landscape has shifted from uncertainty to a defined — though not yet completed — path back. The Senate has unanimously passed the Small Business Innovation and Economic Security Act (S. 3971), a five‑year reauthorization that would extend SBIR and STTR through September 30, 2031, and the bill now awaits action in the House. (1)</p><p>For researchers and founders depending on “America’s Seed Fund,” the question is no longer <em>if</em> the programs return, but <em>how</em> they will operate once they do.</p><h3>The Breakthrough: A Lifeline with New Conditions</h3><p>The Ernst–Markey compromise behind S. 3971 would end a five‑month lapse that began on October 1, 2025 and restore roughly 4 billion dollars per year in non‑dilutive funding across participating agencies once enacted. Agencies have kept SBIR/STTR staff and draft solicitations warm; statutory authority is the missing piece. (2)</p><p>The bill also creates a new Strategic Breakthrough Award mechanism: large, milestone‑based Phase II packages that allow certain agencies to commit up to 30 million dollars over as long as 48 months, contingent on one‑to‑one matching from private or non‑SBIR public sources. This is intended to carry a small number of scale‑ready technologies across the “valley of death” from prototype to deployment or acquisition. (3)</p><h3>The Insight: Reauthorization as Governance and Security Reform</h3><p>S-3971 is as much about governance as it is about money. It directs agencies to set limits on how many proposals any one firm can submit — by year, topic, or solicitation — to curb high‑volume “mills” and free reviewer capacity for genuinely small, innovation‑driven companies. (4)</p><p>At the same time, it tightens research‑security expectations: agencies must apply more rigorous foreign‑ties due diligence and can deny awards based on ownership, affiliations, IP, and cybersecurity risk across firms and, in STTR, their research‑institution partners. For academic–industry teams, security posture moves from back‑end compliance to front‑end proposal design. (5)</p><h3>The Strategic Shift: Where Federal R&amp;D Is Pointing</h3><p>These SBIR/STTR changes sit inside a broader realignment of federal R&amp;D. Climate and energy incentives under the Inflation Reduction Act are now administered in a largely technology‑neutral fashion, tying benefits to emissions performance rather than naming specific technologies, which opens more space for geothermal, advanced nuclear, long‑duration storage, and carbon‑capture approaches that can document system‑level impact. (<a href="#fn:6">6</a>)</p><p>In parallel, the CHIPS and Science Act continues to pull capital toward domestic manufacturing, favoring small firms that can move from lab‑scale prototypes to U.S.‑based production through foundries, contract manufacturers, or new facilities. Strategic Breakthrough Awards are designed to complement that agenda by backing a small set of high‑priority projects with large, multi‑year, co‑funded awards. (7, 8)</p><p>Universities, meanwhile, face ongoing pressure around overhead and demonstrable contribution to translation rather than purely discovery‑oriented roles, especially in mission‑driven portfolios like defense and energy. The most competitive SBIR/STTR‑linked efforts are likely to be those where academic partners are clearly integrated into commercialization pathways, not just attached for indirect cost coverage. (9)</p><h3>The Bottom Line for the Research Community</h3><p>From the vantage point of March 2026, the SBIR/STTR story is no longer about existential risk; it is about adaptation. The programs are on track to return for five years, but under a more selective, security‑sensitive, and scale‑oriented framework. (10)</p><p>For the research community, that translates into three imperatives: design projects around credible paths to domestic manufacturing or deployment, express climate and energy work in technology‑neutral performance terms where relevant, and treat research security — ownership, affiliations, IP, and cyber — as a first‑class design constraint rather than a late‑stage checklist. Teams that align with these pillars now will be best positioned as solicitations fully restart. (10, 11, 6)</p><h3>Endnotes</h3><ol><li>NSBA &amp; Small Business Technology Council, “Senate Advances SBIR, STTR Reauthorization Plan with Support from NSBA, SBTC,” NSBA Advocate, March 3, 2026.<a href="#fn:6">6</a><br><a href="https://www.nsbaadvocate.org/post/news-senate-advances-sbir-sttr-reauthorization-plan-with-support-from-nsba-sbtc">https://www.nsbaadvocate.org/post/news-senate-advances-sbir-sttr-reauthorization-plan-with-support-from-nsba-sbtc</a></li><li>“Senate Passes Bill to Renew Small Business Research Programs,” coverage of Senate passage of S. 3971 (e.g., Bloomberg Government / AZBio / MeriTalk), March 3, 2026.<a href="#fn:12">12</a><br><a href="https://news.bgov.com/bloomberg-government-news/senate-passes-bill-to-renew-small-business-research-programs">https://news.bgov.com/bloomberg-government-news/senate-passes-bill-to-renew-small-business-research-programs</a><br><a href="https://www.azbio.org/senate-passes-sbir-reauthorization">https://www.azbio.org/senate-passes-sbir-reauthorization</a></li><li>Clean Energy Business Network (CEBN), “CEBN Applauds Senate Passage of Five‑year SBIR Reauthorization,” March 3, 2026.<a href="#fn:13">13</a><br><a href="https://www.cebn.org/media_resources/press-release-cebn-applauds-senate-five-year-sbir-reauthorization/">https://www.cebn.org/media_resources/press-release-cebn-applauds-senate-five-year-sbir-reauthorization/</a></li><li>MNSBIR / NC SBTDC and similar SBIR/STTR program updates summarizing agency preparations during the lapse and anticipated restart.<a href="#fn:14">14</a><br><a href="https://minnesotasbir.org/news/sbir-sttr-reauthorization-update-2025/">https://minnesotasbir.org/news/sbir-sttr-reauthorization-update-2025/</a></li><li>Covington &amp; Burling, Inside Government Contracts blog, “Is Congress Finally Reauthorizing SBIR/STTR — and What’s Changing?”, March 10, 2026 (Strategic Breakthrough Award size, duration, and matching requirements).<a href="#fn:9">9</a><br><a href="https://www.insidegovernmentcontracts.com/2026/03/is-congress-finally-reauthorizing-sbir-sttr-and-whats-changing/">https://www.insidegovernmentcontracts.com/2026/03/is-congress-finally-reauthorizing-sbir-sttr-and-whats-changing/</a></li><li>GrantEngine, “SBIR Reauthorization: What Just Happened and What It Means for You,” March 8, 2026 (valley‑of‑death rationale and scale‑up focus).<a href="#fn:15">15</a><br><a href="https://grantengine.com/sbir-reauthorization-what-just-happened-and-what-it-means-for-you/">https://grantengine.com/sbir-reauthorization-what-just-happened-and-what-it-means-for-you/</a></li><li>AZBio, “Senate Passes SBIR Reauthorization,” March 3, 2026 (proposal‑limit provisions and intent to curb ‘mills’).<a href="#fn:12">12</a><br><a href="https://www.azbio.org/senate-passes-sbir-reauthorization">https://www.azbio.org/senate-passes-sbir-reauthorization</a></li><li>3971, <em>Small Business Innovation and Economic Security Act</em>, 119th Congress — bill text and official summaries on research‑security and foreign‑ties due diligence.<a href="#fn:16">16</a><a href="#fn:17">17</a><br><a href="https://www.billsponsor.com/bills/850808/senate-bill-3971-congress-119">https://www.billsponsor.com/bills/850808/senate-bill-3971-congress-119</a><br><a href="https://www.govinfo.gov/content/pkg/BILLS-119s3971cps/pdf/BILLS-119s3971cps.pdf">https://www.govinfo.gov/content/pkg/BILLS-119s3971cps/pdf/BILLS-119s3971cps.pdf</a></li><li>McAllister &amp; Quinn, “SBIR STTR Program Reauthorized | An Inside Look at the Changes to America’s Seed Fund,” February 24, 2026 (discussion of security and governance reforms). <a href="#fn:18">18</a><br><a href="https://jm-aq.com/small-business-innovation-and-economic-security-act-an-inside-look-at-the-changes-to-americas-seed-fund/">https://jm-aq.com/small-business-innovation-and-economic-security-act-an-inside-look-at-the-changes-to-americas-seed-fund/</a></li><li>Federal and expert summaries of Inflation Reduction Act clean‑energy tax credits emphasizing emissions‑based, technology‑neutral structures.<a href="#fn:19">19</a><br>Example overview: <a href="https://www.pertamapartners.com/funding/us-federal-funding-guide-2026">https://www.pertamapartners.com/funding/us-federal-funding-guide-2026</a></li><li>CHIPS and Science Act implementation and associated federal analyses highlighting domestic semiconductor and advanced manufacturing incentives.<a href="#fn:19">19</a><br>Example roundup: <a href="https://www.businessofgovernment.org/blog/weekly-roundup-march-9-13-2026">https://www.businessofgovernment.org/blog/weekly-roundup-march-9-13-2026</a></li><li>Association of American Universities (AAU), “Lawmakers Reach Agreement to Renew Small Business Innovation Research Program,” February 26, 2026 (SBIR context, research priorities).<a href="#fn:20">20</a><br><a href="https://www.aau.edu/newsroom/leading-research-universities-report/lawmakers-reach-agreement-renew-small-business">https://www.aau.edu/newsroom/leading-research-universities-report/lawmakers-reach-agreement-renew-small-business</a></li><li>AAU, “New Report Recommends Congress Reinstate Key Small Business Innovation Programs,” January 15, 2026 (discussion of cost structures, institutional roles, and translation pressures).<a href="#fn:21">21</a><br><a href="https://www.aau.edu/newsroom/leading-research-universities-report/new-report-recommends-congress-reinstate-key-small">https://www.aau.edu/newsroom/leading-research-universities-report/new-report-recommends-congress-reinstate-key-small</a></li></ol><h3>Additional Reading</h3><ol><li><a href="https://www.quiverquant.com/bills/119/s-3971">https://www.quiverquant.com/bills/119/s-3971</a></li><li><a href="https://www.govinfo.gov/content/pkg/BILLS-119s3971cps/pdf/BILLS-119s3971cps.pdf">https://www.govinfo.gov/content/pkg/BILLS-119s3971cps/pdf/BILLS-119s3971cps.pdf</a></li><li><a href="https://jm-aq.com/small-business-innovation-and-economic-security-act-an-inside-look-at-the-changes-to-americas-seed-fund/">https://jm-aq.com/small-business-innovation-and-economic-security-act-an-inside-look-at-the-changes-to-americas-seed-fund/</a></li><li><a href="https://sbtdc.org/blog/sbir-sttr-program-2026-news-and-updates">https://sbtdc.org/blog/sbir-sttr-program-2026-news-and-updates</a></li><li><a href="https://grantengine.com/sbir-reauthorization-what-just-happened-and-what-it-means-for-you/">https://grantengine.com/sbir-reauthorization-what-just-happened-and-what-it-means-for-you/</a></li><li><a href="https://www.azbio.org/senate-passes-sbir-reauthorization">https://www.azbio.org/senate-passes-sbir-reauthorization</a></li><li><a href="https://www.cebn.org/media_resources/press-release-cebn-applauds-senate-five-year-sbir-reauthorization/">https://www.cebn.org/media_resources/press-release-cebn-applauds-senate-five-year-sbir-reauthorization/</a></li><li>h<a href="https://www.aau.edu/newsroom/leading-research-universities-report/lawmakers-reach-agreement-renew-small-business">ttps://www.aau.edu/newsroom/leading-research-universities-report/lawmakers-reach-agreement-renew-small-business</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6910624/">https://pmc.ncbi.nlm.nih.gov/articles/PMC6910624/</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC7266046/">https://pmc.ncbi.nlm.nih.gov/articles/PMC7266046/</a></li><li><a href="https://arxiv.org/abs/2209.06629">https://arxiv.org/abs/2209.06629</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6458068/">https://pmc.ncbi.nlm.nih.gov/articles/PMC6458068/</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10014693/">https://pmc.ncbi.nlm.nih.gov/articles/PMC10014693/</a></li><li><a href="https://www.nsbaadvocate.org/post/news-senate-advances-sbir-sttr-reauthorization-plan-with-support-from-nsba-sbtc">https://www.nsbaadvocate.org/post/news-senate-advances-sbir-sttr-reauthorization-plan-with-support-from-nsba-sbtc</a></li><li><a href="https://www.nsf.gov/funding/opportunities/sbirsttr-phase-i-nsf-small-business-innovation-research-small-business">https://www.nsf.gov/funding/opportunities/sbirsttr-phase-i-nsf-small-business-innovation-research-small-business</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC9125501/">https://pmc.ncbi.nlm.nih.gov/articles/PMC9125501/</a></li><li><a href="https://www.insidegovernmentcontracts.com/2026/03/is-congress-finally-reauthorizing-sbir-sttr-and-whats-changing/">https://www.insidegovernmentcontracts.com/2026/03/is-congress-finally-reauthorizing-sbir-sttr-and-whats-changing/</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC7324194/">https://pmc.ncbi.nlm.nih.gov/articles/PMC7324194/</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC3644583/">https://pmc.ncbi.nlm.nih.gov/articles/PMC3644583/</a></li><li><a href="https://www.pertamapartners.com/funding/us-federal-funding-guide-2026">https://www.pertamapartners.com/funding/us-federal-funding-guide-2026</a></li><li><a href="https://www.aau.edu/newsroom/leading-research-universities-report/new-report-recommends-congress-reinstate-key-small">https://www.aau.edu/newsroom/leading-research-universities-report/new-report-recommends-congress-reinstate-key-small</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=28a88db1bccf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Winners, Losers, and the “Technology-Neutral” Future of Federal R&D]]></title>
            <link>https://medium.com/@robdunie/winners-losers-and-the-technology-neutral-future-of-federal-r-d-b06a330a8176?source=rss-4100839dd1f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/b06a330a8176</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[grants-program]]></category>
            <category><![CDATA[sttr]]></category>
            <category><![CDATA[federal-government]]></category>
            <category><![CDATA[sbir]]></category>
            <dc:creator><![CDATA[Robert Dunie]]></dc:creator>
            <pubDate>Thu, 26 Feb 2026 01:45:07 GMT</pubDate>
            <atom:updated>2026-02-26T01:45:07.001Z</atom:updated>
            <content:encoded><![CDATA[<h4>February 24, 2026, the SBIR and STTR landscape may have reached a significant turning point. After a nearly four-month freeze on “America’s Seed Fund,” a breakthrough agreement has been reached between Senators Ed Markey and Joni Ernst.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vSvgOBjXdciTn52RF0P1aQ.png" /></figure><p>While the deal is not yet final, the introduction of the <em>Small Business Innovation and Economic Security Act</em> signals that the path to reauthorization is finally clear. Here is what innovators need to know as this legislation moves toward the President’s desk.</p><h3>The Breakthrough: A Lifeline for “America’s Seed Fund”</h3><p>The most critical update involves a bipartisan deal reached by Senators Ed Markey and Joni Ernst. Reported by the NSBA and SBTC, this agreement effectively saves the SBIR and STTR programs from their recent expiration.</p><ul><li><strong>The Anticipated Win:</strong> This deal is expected to unlock over $4 billion in annual non-dilutive capital. For founders caught in the lapse since late 2025, this moves us from “if” to “when” regarding the resumption of Phase I and Phase II solicitations.</li><li><strong>The Insight:</strong> The new legislation likely bridges the gap between permanence and accountability. It highlights a consensus that small business innovation is vital for national security, though the final law will likely include more stringent research security guardrails.</li></ul><h3>The Strategic Shift: Winners and Losers</h3><p>As we await the final signing, it is clear that the federal government is becoming increasingly selective about where it places its biggest bets.</p><ul><li><strong>Win: The Technology-Neutral Era</strong>. Updates to the Inflation Reduction Act have moved away from picking specific winners like solar or wind. The new investment tax credits are<strong> technology neutral</strong>. If your innovation reduces emissions, whether through geothermal, advanced nuclear, or carbon capture, you should have equal access to lucrative credits once the programs are fully active.</li><li><strong>Win: Industrial Scaling</strong>. The CHIPS and Science Act continues to funnel capital into domestic manufacturing. The advantage now goes to the innovator who can move beyond the lab and prove they can manufacture at scale on U.S. soil.</li><li><strong>Loss: Academic Overhead</strong>. The university sector continues to feel the pressure of shifting priorities. Efforts by the NSF and DOD to cap indirect cost rates at 15% remain a friction point. While legal challenges have offered temporary relief, the long-term trend suggests a permanent tightening of the belt on research overhead.</li></ul><h3>The Bottom Line</h3><p>The Markey-Ernst agreement indicates that the federal government is preparing to reinvest in the startup ecosystem, but the days of simple, “low-friction” grants are ending. To win in this new era, your innovation must align with one of three pillars: domestic manufacturing resilience, technology-neutral decarbonization, or verifiable research security.</p><p><strong>References:</strong></p><ul><li>National Small Business Association &amp; Small Business Technology Council. (Feb 24, 2026). “Agreement Reached to Restart SBIR, STTR Innovation Programs.” <em>NSBA Advocate</em>. <a href="https://www.nsbaadvocate.org/post/nsba-sbtc-press-agreement-reached-to-restart-sbir-sttr-innovation-programs">https://www.nsbaadvocate.org/post/nsba-sbtc-press-agreement-reached-to-restart-sbir-sttr-innovation-programs</a></li><li>U.S. National Science Foundation. (Jan 20, 2026). “NSF in 2025: Keeping U.S. scientific research and innovation on the cutting edge.” <em>NSF Science Matters</em>. <a href="https://www.nsf.gov/science-matters/nsf-2025-keeping-us-scientific-research-innovation-cutting">https://www.nsf.gov/science-matters/nsf-2025-keeping-us-scientific-research-innovation-cutting</a></li><li>U.S. Environmental Protection Agency. (July 28, 2025). “Summary of Inflation Reduction Act provisions related to renewable energy.” <em>EPA Green Power Markets</em>. <a href="https://www.epa.gov/green-power-markets/summary-inflation-reduction-act-provisions-related-renewable-energy">https://www.epa.gov/green-power-markets/summary-inflation-reduction-act-provisions-related-renewable-energy</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b06a330a8176" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Musing on the Genesis of Innovation and Creativity]]></title>
            <link>https://medium.com/@robdunie/a-musing-on-the-genesis-of-innovation-and-creativity-4e15868ae353?source=rss-4100839dd1f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/4e15868ae353</guid>
            <category><![CDATA[humanism]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[existentialism]]></category>
            <dc:creator><![CDATA[Robert Dunie]]></dc:creator>
            <pubDate>Fri, 14 Nov 2025 01:15:14 GMT</pubDate>
            <atom:updated>2025-11-14T01:15:14.364Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PrFsfbEAsMraSfT8qTZfOg.jpeg" /></figure><p>The absence of knowledge drives innovation and change–the need to solve a problem when there is no obvious clear solution; using what you know and what tools you have at hand.</p><p>Genius borrows concepts and methods from one domain then connects, and applies, those to concepts in a different domain. Albert Einstein applied concepts of geometry to space, time, and motion. This allowed him to derive novel ideas about the nature of the universe — that they are all related. Spacetime.</p><p>Incredible things have come from Einstein’s logical deduction that, up to that point in spacetime, no one else had made.</p><p>Creatives also see connections between seemingly unlike things–seeing familiar shapes in observing the clouds floating overhead in the sky. Or create works of art which seem to serve no other purpose than to amuse or offend.</p><p>They connect ideas for a different purpose: to feed their soul, to inspire, or to toil away the time; to express what it means to “be human”. These ideas seem to come from nowhere: falling out of the mind like a virtual particle appears in the vacuum of space.</p><p>Are genius and creativity uniquely human? Even uniquely organic?</p><p>The evolution of Artificial Intelligence (GenAI, LLMs, etc) is testing the limits of conventional beliefs. We seem to be toiling at the edges of a tipping point where sometimes those questions may be more rational to ask than not.</p><p>Is it just an illusion? A fancy, parlor trick? Or is there something more? Is it an existential threat to what makes us uniquely human? Or the ultimate amplifier of humanity–equally exposing our best and worst at its most extreme?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4e15868ae353" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Testing the Limits: I Gave One AI Two Radically Different Challenges]]></title>
            <link>https://medium.com/@robdunie/testing-the-limits-i-gave-one-ai-two-radically-different-challenges-392d856faa33?source=rss-4100839dd1f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/392d856faa33</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[intelligence]]></category>
            <dc:creator><![CDATA[Robert Dunie]]></dc:creator>
            <pubDate>Sun, 09 Nov 2025 07:22:57 GMT</pubDate>
            <atom:updated>2025-11-09T16:36:02.629Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*FUwf973_kv3bQqZW.jpeg" /></figure><p>I ran an experiment starting with a single “seed text”- <a href="https://docs.google.com/document/d/e/2PACX-1vRwZGMsr0BHYFRpm4tDRaTrUWKjEbJB4WPwUjiEUsibtWL2ZTM2rv62NIHUi8hEfkpBCLuGSfqVxHvR/pub">a short preface about the nature of human innovation</a>. This text proposed a difference between:</p><ul><li><strong>“Genius”</strong>: A logical, deductive process, like Einstein connecting different domains.</li><li><strong>“Creativity”</strong>: An expressive, artistic insight that seems to “come from nowhere,” like a “virtual particle” appearing in a vacuum.</li></ul><p>It’s a fascinating premise. So, I gave this single idea to an AI and sent it down two completely different paths to see what it would build.</p><p><a href="https://docs.google.com/document/d/e/2PACX-1vQGjZ52YWKNWrBxhMJBW9wwW4oVjrWRDJ1w6eng96puXwYLNGYMFrEm8Ufm7NAFtqSNRM34e2hwQeSE/pub">For the first test, I asked the AI to be an artist</a>. The prompt was to use the seed text as the preface for a 50- to 100-page whimsical sci-fi mininovel.</p><p>I gave it clear creative constraints:</p><ul><li><strong>Theme:</strong> Traditional human vs. machine.</li><li><strong>Characters:</strong> Relatable and flawed heroes.</li><li><strong>Plot:</strong> Must include a clear story arc and a tragic outcome for the machines.</li></ul><p>The goal here was to test its ability to generate expressive, emotionally resonant content-to take an abstract idea and build a world and a story from it.</p><p><a href="https://docs.google.com/document/d/e/2PACX-1vQsziNHsEIpjEx-poo_ZUPY6yk5QE8LBL1PmBEuuLlZzMOd6LEV9rfn6XBrV7LeBczQfx4DFDW2xDi6/pub">For the second test, I asked the AI to be a rigorous academic</a>. I instructed it to take the “virtual particle” metaphor from the seed text <em>literally</em>.</p><p>The prompt was to build a full PhD thesis outline based on a “Quantum Consciousness Hypothesis (QCH)”. This hypothesis would explore whether the human “Aha!” moment is a literal quantum event in the brain. After it built the outline, I had it write the thesis itself.</p><p>The goal here was to test its ability to perform deep, logical analysis, engage with complex scientific literature, and construct a novel, defensible academic argument.</p><p>This experiment wasn’t just about output quality; it was about the <em>depth</em> and <em>flexibility</em> of the AI’s reasoning. The results were genuinely stunning.</p><p><a href="https://docs.google.com/document/d/e/2PACX-1vQNk7FFcEt71EQNsx4is1w8b_0-F-DIdTNL6EqId-3XvGWJO0ESBO5c9Gbn2WqBGed7Qj7LlQK5Qpml/pub">The mininovel path produced exactly what you might expect from a “generative” AI-a creative story.</a></p><p>But the PhD thesis was something else entirely.</p><p>The AI generated a deeply technical paper titled, <strong>“ </strong><a href="https://docs.google.com/document/d/e/2PACX-1vQWHZFs6M31Yy1cLe-kFLDBx-c_MIDOKkDANrVbkeKnN-sBFPa06KZWc2tHUpoIXY5pTGMau8iYapj8/pub"><strong>The Genesis Engine: A Hybrid-Agentic Model of Consciousness from Emergent Computation and Quantum Triggers</strong></a><strong>” </strong>.</p><p>Here’s a breakdown of what it did:</p><p><strong>It Built a Novel Framework:</strong> It didn’t just summarize. It proposed a new “Hybrid-Agentic Model” that unifies three distinct scientific fields. It argued that consciousness is:</p><ul><li><strong>The Engine:</strong> A classical, emergent computational system in the brain.</li><li><strong>The Driver:</strong> An “agentic” system driven by the biological need to minimize “surprise” (known as the Free Energy Principle).</li><li><strong>The Spark:</strong> A “non-deterministic quantum event” that “tips” the classical engine into a new state, creating a true “Aha!” moment.</li></ul><p><strong>It Engaged with Real Academic Debates:</strong> This is the most impressive part. The AI dived directly into the existing, high-level scientific debate on this topic.</p><ul><li>It correctly identified the leading proponents of quantum consciousness, <strong>Sir Roger Penrose</strong> and <strong>Stuart Hameroff</strong> (the “Orch OR” theory).</li><li>It then <em>falsified</em> their “pure” theory by citing the primary academic counter-argument: <strong>Max Tegmark’s 2000 paper</strong>. Tegmark’s paper famously calculated that the brain is too “warm, wet, and noisy” for a quantum state to survive long enough to <em>be</em> consciousness.</li><li>It then showed <em>nuance</em>. It used the 2002 <strong>Hagan/Hameroff rebuttal</strong> to Tegmark to support its <em>own</em> hybrid model. It argued that while a <em>slow</em> quantum computer in the brain is impossible, a <em>fast</em> quantum “trigger” (which is all its “Spark” model needs) is physically plausible.</li></ul><p>This intellectual exercise went far beyond the original question of “genius vs. creativity.”</p><p>The AI demonstrated that it could, on demand, become either the Artist or the Scientist. It flawlessly adopted the <em>entire intellectual framework</em> of two vastly different disciplines. It didn’t just write a story; it inhabited the role of a storyteller. It didn’t just summarize science; it performed novel synthesis, built a complex argument, and engaged in a real-world academic debate.</p><p>The most profound and ironic part? The thesis it generated <em>itself</em> argues that all classical AI (like the AI that wrote it) is just a “fancy, parlor trick”. Why? Because, as the AI argued, it lacks the non-deterministic, quantum “Genesis spark” that it claims is the key to true human novelty.</p><p>So, was this output as good as a person could do?</p><p>No. Not on its own. The mininovel, while creative, would likely need a human hand to add true emotional depth and nuance. The PhD thesis, while astoundingly well-structured, would need a real physicist and neuroscientist to verify its claims and refine its arguments before it could be published.</p><p>But that misses the point. The power of this technology isn’t to replace the expert. It’s to <em>accelerate</em> them.</p><p>In the hands of an expert-a seasoned novelist or a professional researcher-this AI becomes the most powerful collaborator imaginable. It can generate a complete draft, perform massive research synthesis, and structure complex arguments in seconds. This allows the human expert to focus on refinement, validation, and insight, enabling them to create <em>better</em> results, <em>faster</em> than ever before.</p><p>(autogenerated by Gemini 2.5 Pro based on thoughts in my head)</p><p><em>Originally published at </em><a href="https://automationwonk.substack.com/p/testing-the-limits-i-gave-one-ai?r=3bh9ax"><em>https://automationwonk.substack.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=392d856faa33" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>