I wrote the contract. Agents keep the research alive.

The situation

I keep a public knowledge base on the behavioral layer of AI: how agents express confidence, handle uncertainty, escalate, and recover when they're wrong. It is a working research practice, and a research practice is only worth anything if it is accurate and current. Both are expensive to maintain by hand, and they decay the moment you stop.

The obvious move is to let AI agents write and publish it. That is also the exact failure my whole body of work is about. An agent that publishes a plausible but wrong claim does not just make an error; it poisons the credibility of everything around it. Accuracy is the brand. For a site whose entire subject is whether an AI can be trusted, an untrustworthy pipeline would be self-refuting.

The reframe

So I asked the question I ask of any agent product. Not whether to use AI, but where the human belongs. The failure mode here is not the drafting, it is the publishing. A draft is recoverable. A wrong claim shipped to a public site under my name is not. The irreversible step is the merge.

That located the whole design. Put the human on the irreversible step, and govern everything before it with a standard precise enough to enforce. Agents propose; I dispose. The human merge is the point of the entire thing, not a formality bolted on at the end.

The work

The standard is a written editorial contract. It is numbered and enforceable, so a person or an agent can follow it without further context. Every published note must cite a tier-one primary source, or it fails the build. Every note carries both a builder read and an exec read, the same dual audience discipline as my book. Freshness is tracked: a note that has not been re-verified in thirty days is surfaced for review. House style and link integrity are checked. The clauses that can be machine-checked are wired into continuous integration, so a pull request that violates one cannot merge.

The most important clause is the one about what the machine cannot do. CI verifies structure, required fields, links, and a confidential-term denylist. It cannot verify that a claim is actually true, or that a source is genuinely primary. That judgment stays the author's, and the contract says so in as many words: it is the most important rule in the document. This is the same split I build everywhere. Encode what can be checked and enforce it; keep the part that needs judgment a human's job, on purpose.

Under that contract, agents do the drafting. One skill researches and drafts a note on demand. Another runs a weekly cadence: it scans the past week for concrete, primary-source events in the field, verifies each against the source it actually read, and drafts a small, tight batch. Both do the identical thing at the end. They run the contract's checks, and they open a pull request. Neither ever merges. When a candidate brushes against work I cannot discuss, the agent is instructed to leave it out and flag it for me rather than publish it, and a denylist backs that up in CI.

There is a deliberate reflexivity to it. The site is about the behavioral layer, and it is built with one: a behavioral contract governing a set of agents, a human placed at the consequential step, and confidence, escalation, and repair designed into the workflow rather than hoped for. The method and the subject are the same thing. That is not a gimmick. It is the strongest form of the argument, because the site cannot make a claim its own construction contradicts.

The outcome

It runs. A steady trickle of primary-source notes and a weekly briefing outline arrive as pull requests; I review and merge; the merge deploys. The site is live, and the entire trail is public. You can read the contract, read the agent skills that draft under it, and read every pull request that built the site, each one listing its sources and the checks it passed.

That is the thing no slide can do. The proof is not that I say I can operationalize AI across a craft. It is that you can open the commit history and check. And it is the same method as the governance system I run at work, now in a second domain: there, a contract governs how a product is designed; here, a contract governs how a research practice is run. Two domains, one method. The difference is that this one is entirely mine and entirely public, so there is nothing to take on trust.