DoorDash has built a multi-agent workflow to remove stale feature flags from a codebase containing more than 60,000 flags across roughly 623 repositories. In a 50-flag evaluation, the system produced 45 usable pull requests in an average of 13.8 minutes and at an average cost of $4.79 per cleanup. DoorDash estimates manual removal takes one to two hours.
An orchestrator using Claude Sonnet gathers Jira tickets, searches repositories and retrieves live rollout data through the Model Context Protocol. An engineer reviews that report and confirms the value that should remain before any source code changes begin. Claude Opus agents then work in separate Git worktrees, allowing up to four parallel cleanups without sharing files or build state.
Each agent updates code and tests, then runs builds, test suites, patch-coverage checks and static analysis. A pull request opens only after those checks pass. Of the 50 trials, 31 merged on the first attempt, 14 needed revisions and five required engineer intervention, mainly for deep call chains and parameters threaded across interfaces. DoorDash reported no bugs or regressions in the evaluated changes. The company plans to add confidence scores for low-risk jobs and a final code-quality pass, but the current design deliberately keeps a human approval gate before agents edit the code.