Stop re-deriving what someone already solved
Six waves of work produced a real autonomous execution loop and zero pages of website. The correction did not come from building more of it — it came from reading what the company next door had already built, run for four days, and deleted.
Text-first. The recording is pending the first real asset through the Drive intake adapter; this page does not claim a video exists.
Key moments
- The queue that never had an input
- Zentari deleted the thing we were building
- The runner is the lease
Transcript
The instruction that reframed everything was four words long: stop re-deriving what Zentari already solved.
The engine had a left half and a right half. The left half ran — intake refilled its own queue, the loop fired on a schedule, receipts recorded intent before effect. The right half had never had an input. Zero assets ingested, zero learnings, zero published artifacts, and a website with no pages on it.
Next door, the same machinery had already been built and thrown away. A Fargate drain pool with circuit breakers and snapshot state produced zero autonomous merged pull requests across about four days. A single standalone GitHub Action, one run per unit of work, produced a real merged pull request in about five minutes.
So the leasing, the heartbeats, the file lock, the reclaimable expiry — all of it was careful work solving a problem that only exists if you insist on one long-lived worker. One runner per packet, and the runner is the lease. Nothing to strand, because nothing persists between runs.
The part worth keeping is smaller and less impressive than the part worth deleting, and that is usually the shape of it.
Published 2026-07-30