What are you building with agents outside the chat window?

Three, all on a launchd heartbeat every 15 minutes, none in a chat window. A code reviewer that reads the diff since its last run and must end with a parsed verdict line. A docs agent that edits only Markdown inside a throwaway git worktree, with a guard that discards the whole patch if it touched code. A QA agent that boots the app, reads the live DOM, picks the next click, and files a bug report with screenshots.

What went wrong first: the docs guard let code files through as "docs" on the first adversarial pass, so it is now a denylist on top of the allow-list. The QA agent, given a soft "please wrap up", ran 103 steps and $0.67 on one flow; it now has a hard step budget and the action tools refuse past it. And the cheap QA engine found eight cosmetic bugs and missed the one that mattered; the version with session memory found a change-photo-then-re-search flow that froze the app forever.

The rule I landed on: nothing merges on its own. Every agent's output is a patch or a report a human reads.

Trusted one with code review on changes that alter a reported number. What it got wrong first: a change read the total at one block and the flows at a later block. Spec said so, code matched, tests matched, review said conformant. Anything landing in the gap between the two blocks was counted wrong. A human wrote the failing test.

The lesson was not agents bad. Spec, code, tests and reviewer all came out of one head, so they agreed with each other and not with the chain. Rule now: state the invariant in one sentence, and test the case where the two clocks differ, not the case where they happen to agree.

Lazy question back: your morning transcript reader, does it read for what the agent said, or for what it did not do? 🕰️