I wrote the contract. Agents draft under it. I merge.

The situation

I keep a public knowledge base on how AI 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 the site is about. An agent that publishes a plausible but wrong claim doesn't just make an error; it poisons the credibility of everything around it. For a site whose 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 isn't the drafting. It's the publishing. A draft is recoverable. A wrong claim shipped under my name costs something a revert doesn't get back. The consequential step is the merge. Put the human there, and govern everything before it with a standard precise enough to check. Agents propose; I dispose.

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 include a source citation. CI checks the required fields and links; I assess whether the source is primary and supports the claim. Every note carries both a builder read and an exec read. Freshness is tracked: a note that hasn't been re-verified in thirty days is surfaced for review. House style and link integrity are checked. The clauses that can be machine-checked run in continuous integration on every pull request, and I don't merge one that fails.

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: it is the most important rule in the document.

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 read, and drafts a small batch. Both run the contract's checks and open a pull request. Neither ever merges. When a candidate brushes against work I can't discuss, the agent is instructed to leave it out and flag it for me, 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 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. Its construction can be checked against its claims.

The outcome

It runs. 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 whole trail is public: the editorial contract, the agent skills that draft under it, and every pull request, each listing its sources and the checks it passed.

This is my own operation, one person and a set of agents, and it's a different thing from leading a team. What it shares with the harness at work is the method: a written standard, the checkable part checked, a person on the step that matters. Here the whole of it is public.