DoorDash describes a narrow coding task with explicit safety checks: remove obsolete feature flags, the settings used for gradual releases and experiments. The system queries deployed values before an engineer approves each proposed replacement.
- Of 50 Kotlin cleanups, engineers accepted 31 immediately and 14 after minor revisions. Five required substantial intervention. The 90% usable-PR rate includes revisions.
- Each removal agent edits in an isolated Git worktree. Builds, tests, linting and 95% coverage of changed lines check its work.
- CI caught six initial coverage failures that the model’s own assessment missed.
- Mean runtime was 13.8 minutes, with $4.79 in model API charges per cleanup. The manual one-to-two-hour comparison comes from internal estimates.
The transferable design is authoritative live state, a human decision point and checks outside the model. The company’s 56,000 flags describe its environment; only 50 removals were evaluated.
Caveat: This is an author-reported study of closed company repositories, with each flag processed once. Repeatability and transfer to other codebases remain open.