DoorDash built a production system in which a team of coordinated AI agents finds, analyzes, and removes stale feature flags from its code. Feature flags are on/off switches engineers leave in code to control whether a feature is live. The system handled 35,000 changes at a 90% approval rate, with each job completing in under 14 minutes at a cost of $4.79. Engineers reviewed proposals before any code changed, but the agents did the work of locating affected files, reasoning about code behavior, drafting the change, and writing the change summary. The architecture assigned distinct roles to distinct agents: one scanned repositories for flag usage, one reasoned about whether removal was safe, one drafted the code change, and one wrote the change description for the reviewer. Each agent received a bounded task and passed its output to the next. The engineers who approved or rejected the final change saw a structured summary the agents composed, not raw agent logs. The strongest design signal here is the approval step. Reviewers made a go or no-go call on a change that touched production code, armed with a summary the agents wrote. That summary had to be trustworthy enough to act on. The 90% approval rate suggests it was reliable, but the 10% rejection rate is the interesting number: it tells you the reviewer was doing real judgment, not rubber-stamping. For teams designing similar systems, the quality of that handoff summary is the variable most worth instrumenting.