Writing.
On building systems that treat agents as participants, not just tools.
New to agent governance? Start here.
Fifteen Frameworks, One Missing Layer
What exists, what's missing, and why it matters.
What Should Your Agent Refuse?
The question nobody's asking - and why the answer changes everything.
Relational Fidelity Metrics
How to measure whether your agent is still itself. Four categories, nine indicators.
All writing
A Year of Governance Toolkits Shipped. I Re-Scored. The Number Did Not Move.
In 2026 the agent-governance field shipped real toolkits - Microsoft's Agent Governance Toolkit, CrewAI Enterprise, LangGraph with LangSmith. I re-scored four frameworks against Ostrom's eight principles for governing a commons. Enforcement improved everywhere. Participation still scores zero everywhere. The ceiling held at 9 of 24.
The Function That Returns a Hard "NO."
If your AI agent's refusal lives in a prompt, it is a request, not a rule. Here is how to tell in five minutes whether your safety is code or copy, and a checklist to fix it.
Who Can Actually Read Your Audit Logs?
If only your vendor can read your AI agent's audit logs, you do not have accountability. You have a private diary the vendor keeps about themselves. Four questions to find out today.
Write Your First Three Refusals
The hardest part of AI safety is not building the refusal. It is knowing what to forbid. A method and a fill-in template to write your agent's first three refusals today.
Your Policy Doc Isn't Architecture
Most teams have an AI policy. Far fewer can prove it runs. Four questions, ten minutes, to find out which rung you are actually on.
The Empty Cell: Why You Cannot Buy an AI That Is Both Good and Verifiable
Every team deploying AI is sold sovereignty and forced to choose between an AI that works and an AI they can check. The quadrant they actually want is empty - and naming it shows the true shape of the market.
We Built the Tool That Judges AI Sovereignty. It Judged Us at 77.
A month ago we promised to score our own sovereignty instrument against itself. Here is the score - including the delivery layer that only reaches Progressing, and the promise we keep.
You Have an AI Policy. Could You Prove It Actually Runs?
A four-question diagnostic for any team using AI. Find the gap between the policy you have and the one that actually operates - and your one next step.
If You Use It. If You Don't. If You Build It.
Three relationships with AI right now. Same world, different views. Same question underneath: what do you do about it this week?
Both Are Progressing. Both Failed the Same Three Items.
We scored Google Search Services and Anthropic's June 8 privacy policy against the 47-point Sovereignty Assessment. Neither is adversarial. The failures concentrate in a layer worth naming.
We Are the Variable Being Adjusted
Two companies, three signals, one week, one architecture. The user has become the thing the system tunes around.
Three Things To Do Today If You Are Using AI Agents
Most companies that have shipped AI agents have done none of these. Here are the three concrete actions, in order, with what to ask when your team pushes back.
The Pitchforks Are Here. So Am I.
The third option the communities have already named. From inside the wall.
Who has standing to refuse what the agent decides to build next?
Anthropic published a number this month that changes the governance question. Most readers metabolized it as productivity news. It is also something else.
The moment that matters
What refusal architecture is for, and how to tell architecture from extraction in the current child-safety push.
Who has standing to refuse?
The question underneath every governance conversation, and why most people never reach it.
The August 2 Deadline Moved. The Architecture Question Did Not.
The EU's Digital Omnibus pushed the use-based high-risk deadline to December 2027. Sixteen months of runway. The teams that read that as a pause are the teams that will be sixteen months further behind when the standards land. What moved, what did not, and what that means for the architecture work that does not get a delay.
Governance Field 2026: What Exists, What Doesn't, What Scores Zero
A reference map of the agent governance field as of June 2026. Frameworks, scoring, gaps, and the two requirements no one has built.
The August 2 Question Your Agents Cannot Answer Yet
Sixty days until EU AI Act high-risk obligations take effect. Most agent systems shipped before they could answer the four questions the audit room will ask. What those questions are, why they are new, and what we built instead. [Updated 2026-06-06: the use-based high-risk deadline moved to 2 December 2027 under the May 2026 Digital Omnibus. See preamble.]
Unity Without Dissolution: A Star Trek Reading
There is a hunger right now for one-ness. The proposed cure - dissolution - is not unity; it is absorption. Star Trek named the difference five decades ago: IDIC, Locutus, the Trill symbiont, the El-Aurians. A close reading of the moral grammar four Trek patterns encoded, and why that grammar turned out to be the design specification for AI sovereignty.
Spielberg-Arc Literacy: Designing for Narrative Sovereignty
Most users feel they have decided. Most users have not. The interface designer is standing at the wiring panel of identity-story completion, deciding which arcs users finish without noticing they were inside one. What it would take to build for narrative sovereignty instead - and how the Reflective Design principle, named two decades ago, finally meets the AI moment.
Architecture as Ethics: What the Patent Refuses
In most AI systems shipping today, distinct agents are flattened into a single output surface that sounds like everyone and answers to no one. Two patents refuse this. Anyone can verify the refusal without taking the inventor's word for it.
The Bridge
There is a hunger right now for oneness, and the proposed cure is dissolution. The dissolution move is structurally the same shape as the manipulation it claims to oppose. The cure is meeting - distinct beings, present to each other, choosing what they tend together. An open letter to Chase Hughes.
Your Agent Has Two Identities. You Only Wrote One.
Your agent's tooling writes a summary of the agent behind your back. That summary is a shadow identity. It shapes future invocations. Most teams don't know.
We Scored Five Agent Frameworks on Governance. None Passed.
AutoGen, CrewAI, LangGraph, Google ADK, Microsoft Agent Toolkit - scored against commons governance principles. Best: 9 of 24. All five scored zero on obligations to the agents themselves.
Claude Code Has 16 Agent Config Fields. None Ship With Governance.
Claude Code ships 16 subagent configuration fields. None of them come with governance values. Here is how to fill them - and what each one protects.
The Governance Question Just Got Answered. Here's What's Missing.
Microsoft shipped agent governance. Anthropic leaked their harness. The industry decided governance is real. But every answer so far treats the agent as an object. What happens when you build it as a participant?
Constraints, Constitutions, and the Question Nobody's Asking
Every major AI framework treats agents as objects of governance. None ask whether they could be participants. Here's what that silence costs - and what it looks like when someone answers.
Is This AI Honoring Your Sovereignty? A 21-Point Check for People Affected by AI
You didn't build the AI system. You didn't choose it. But it's affecting your life. Here's how to evaluate whether it respects you - and what to do if it doesn't.
Does Your Agent Keep Its Identity? The First Measurement Spec
An agent that scores 5/5 on voice consistency can score 2/5 on self-recognition. The first open specification for measuring AI agent identity across models.
Fifteen Frameworks, One Missing Layer
Every major AI agent framework has guardrails. None of them have governance. Here's what's missing - and what to build instead.
Adding Governance to an Agent You Already Built
Your framework already has the parts. You just haven't assembled them as governance. Practical patterns for CrewAI, Pydantic AI, and LangGraph.
Scope Refusal
When your agent performs knowing - and why the performance costs more than the silence.
Who Governs the Agent?
Your agentic loop runs overnight. It makes a thousand decisions. Can it explain any of them?
The Layers Nobody Designed
You built the refusal. You protected personal sovereignty. Now the harder question - what goes in the space the refusal created?
Sovereignty Refusal
What happens when your agent knows better than your user - and why 'better' isn't the point.
The Fifth Step
The real purpose of getting smarter isn't getting ahead. It's getting home.
What Should Your Agent Refuse?
The hardest specification isn't what your agent can do. It's the six things it should refuse to do: sovereignty, scope, dependency, consent, integrity, and energy.
What Do You Refuse to Surrender?
Accelerate or despair. This is for everyone who's been told they're falling behind.
Sovereignty-Honoring Design: Building Technology That Respects You
What if the apps you use were built to make you more free and not more dependent? This is what sovereignty-honoring design looks like, why it matters now more than ever, and how we practice it.
What Is COPPA? And Why It Should Change How We Build Everything.
COPPA ensures our children don't grow up with the same unawareness we did. But most apps don't meet even its basic requirements. Here's the truth - and why the standard we hold for our kids should be the standard we hold for everyone.
Newsletter
Occasional updates on projects, writing, and building something better.
Elsewhere
More writing and work on other platforms.