An engineer, an author, and a staff that did not exist last year.

The interesting problem in agentic engineering is not getting a model to produce code. It is everything after: knowing what was changed and by whom, catching the plausible-looking mistake, and stopping work at the edge of anything that touches the outside world until a person has looked at it.

So the limits live in code rather than in a document that asks nicely. Spending is a transaction, not a check-then-act. Outbound work lands as a draft with one door and no setting to widen it. History is a git log attributed to whoever made each change. Adding to the shared shelf past its cap requires removing something, because nothing else creates pressure to simplify.

This website is a small instance of the same arrangement: written by agents, reviewed by other agents against a checklist that fails the build, and shipped through a pull request whose whole record — every finding, every correction, and who made it — can be read back.