method
How I work
I operate crews of AI agents the way a company operates engineers: isolated workers, a visible board, a supervisor who can watch, speak to and stop each one, and an operating record where every session ends in a verdict.
Six mechanisms, each one a thing in the repository rather than a principle on a slide. They exist because agent work fails in a specific way: it looks finished long before it is checked.
Three of the systems this method runs on are written up as case studies — the method applied rather than described. The full register of seven is on the front page.
six mechanisms
What the structure actually is
Each row names a mechanism and the failure it exists to prevent. The evidence is not on this page — it is on the page each row links to.
where to check
The mechanism leaves a trace
A method that cannot be inspected is a preference. Each of these is a page on this site, rendered from its own source — none of them asks you to take the claim on faith.
What this is not
Not a claim that the agents run themselves. The autonomy is in the workers; the discipline is in the structure around them. A crew without a supervisor is an outage with a schedule.
Not a claim of scale I have not measured. Where a number appears on this site it is counted from a recorded source, and the source is named next to it.
Not a claim that the verdict convention is mechanised end to end. The gates and the freeze are scripts; the reporting contract is a convention with a merge gate behind it.
Not a complete record: 4 published post-mortems, and a board that still lists open items — including the ones nobody has closed yet.