DoorDash entrusts the household of its 1,659 features obsolete flags to LLM agents: 90% of PR accepted, for 4.79 dollars per piece

Out of 50 obsolete flags, the multi-agent system produced 45 requests validated by engineers, in an average of 13.8 minutes.

At DoorDash, more than 56,000 feature flags, the switches that allow a feature to be gradually activated, are spread across 623 code repositories. By May 2026, 1,659 of them had become useless: once the feature is deployed everywhere, the flag and the code it conditions become technical debt, or even a risk if a dormant path reactivates by mistake. Internally, each manual cleaning is estimated to take one or two hours, and rule-based tools were not enough.

Two engineers describe the system they built on arXiv: LLM agents query the actual deployment data via MCP, read the associated Jira ticket, modify the code, and open the end-to-end pull request. Evaluated on 50 stale flags, it resulted in 45 engineer-approved PRs, or 90%, at an average cost of $4.79 and 13.8 minutes per cleanup.

The article also details the limitations encountered and the architectural lessons learned from the project, for teams tempted to deploy this type of agent on a large codebase.