Field notes · No. 011
Senior DevOps Engineer building in public. Shipped features across Local Fitness, Ghostwriter, and more — the tradeoffs behind them, and what broke.
An autonomous review loop should not spend the same budget on a typo and a production change. Persist a complexity profile once, use it to choose the review depth, and stop with the current artifact when the budget runs out.
Keep a full pull-request patch for context and a second patch for the fix since the previous review, then size the next review around that delta.
Running one agent session per open issue against one checkout works until two sessions pick the same issue. A claim in two places, an exclusive create as the tiebreaker, a marker that lets the claim expire, and one forced fetch before every branch.
An agent that can split work into stacked pull requests will split whenever the rule allows it. Make one pull request the default and make the parser refuse a split that does not state, in numbers, what makes the change too large to review as one.
Build a bounded review controller with complete artifact publication, run-and-revision matching, file-level rechecks and a blocked exit when major findings remain.
If a machine is going to approve work instead of a person, the approval has to be bound to the exact bytes it reviewed. Build a registrar that hashes the artifact, records the commit, and refuses to advance when either has moved.
Show elapsed time, observed file activity and historical durations without mistaking an mtime for successful delivery or worker liveness.
A gate that parses test-runner output worked in CI and refused every real run on my laptop, because Node changed its default reporter out from under it. The fix is a parser that accepts every summary the runner actually emits.
Ordering steps by their position in a list is not the same as ordering them by what they need. One measured run spent an hour at a gate enforcing a dependency that did not exist.
A staged agent pipeline where the rule that stops a stage running early is a function, not a sentence in a prompt. Here is how to build the gate, and the four refusals that make it worth having.
