DoorDash's multi-agent system turns stale feature flags into merge-ready PRs for $4.79 and 14 minutes each
DoorDash manages more than 60,000 feature flags ("dynamic values") across about 623 repositories, with over a thousand stale and about 2,300 new flags a month; a manual cleanup takes one to two hours. It built a two-phase system on Google's Agent Development Kit: a Claude Sonnet orchestrator pulls stale-flag Jira tickets, reads live rollout state from the experimentation platform over MCP and proposes a target value, an engineer confirms it, then Claude Opus removal agents run in parallel git worktrees (up to four per repo, one-hour timeout) and may only open a PR once build, tests, 95% patch coverage (JaCoCo) and Detekt pass. On the 50 most recent stale flags in Kotlin repos it produced usable PRs for 45: 31 merged first shot, 14 after one minor revision, five needed an engineer. The work was accepted for the ICSME 2026 industry track.
The numbers (as reported)
| Usable PRs | 45 of 50 stale flags |
|---|---|
| Merged first shot | 31 of 50 |
| Average cost per flag | $4.79 |
| Average time per flag | 13.8 minutes (vs 1-2 hours manually) |
| Bugs or regressions introduced | None observed across 50 PRs |
| Scale | 60,000+ flags across ~623 repositories |
Facts
- Company / subject
- DoorDash
- Tools
- Claude
- Models
- Claude Sonnet, Claude Opus
- Codebase
- brownfield
- Work type
- Refactor / maintenance
- Stack
- Kotlin, Gradle, JaCoCo, Detekt, Jira, MCP, Google Agent Development Kit
- Published
- 2026-08-24
- Last checked
- 2026-11-06
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- Usable PRs for 45 of the 50 most recent stale flags 45/50
“the system produced usable PRs for 45 of them at an average cost of $4.79 and 13.8 minutes per flag”
- 13.8 minutes per cleanup instead of one to two hours manually 13.8 minutes
- No bugs or regressions introduced across all 50 generated PRs
“The system introduced no bugs or regressions along the way.”
- Using Sonnet for orchestration and Opus only for editing cut overall API cost
Caveats
Self-reported by the team that built it, on 50 flags in Kotlin repositories. A human still confirms each target value and reviews the PR. Complex flags reached 85% (22/26) and outliers took about 40 minutes and $19.
More stories by tool: Claude
Related topics: Refactoring and maintenanceBrownfield codebases