The delivery crewcrew
Go on. Break CI.
Agents pick up work, implement, review, run QA and report across three model runtimes. A sweep with no model in it decides who wakes up and when.
sweep / 10 min, no LLM SHA moved? wake one role implement / review / QA red 3 heads? driver parks or merges
Constraint
One person’s attention, three model runtimes, a forge that never stops changing.
What it refuses
- Waking anyone when the commit did not move. A plain-code sweep runs every 10 minutes and tracks the reviewed SHA.
- A second wake for the same head. Red CI wakes the implementer once.
- A fourth retry. The same check red on 3 heads goes to the driver, who parks or re-plans.
Outcome
57 of 86 issues closed by the driver under a delivery-speed bar. Time per ticket and cost per PR: not measured yet.
What broke
- Four shallow approvals landed on a PR that did not compile. Now nothing with a red or pending check gets approved, and 41 earlier approvals were dismissed.
- Batching the reviewer hourly saved money and dropped 4 of 7 real misses. It went back to the 10-minute sweep with review lenses.