DoorDash's AI agent ran 130,000 engineering tasks in its first month

DoorDash just quietly stood up its own AI agent platform. The delivery company's internal system, called Flux, has already run 130,000 engineering tasks in its first month — and it's now cranking out more than 25,000 code reviews a week.

Flux works by spinning up a Firecracker microVM for every task, loading the relevant codebase, dev tools, and permissions on demand. To keep agents from wandering off inside DoorDash's internal systems, the platform routes through a custom MCP gateway. Each task can only call tools and permissions that were declared in advance, and every action is logged for audit.

Engineers can package tasks as Playbooks and fire them off from Slack, GitHub, the CLI, or scheduled jobs. That's the next step beyond the AI code generation DoorDash has already leaned into hard — CEO Tony Xu says nearly two-thirds of the company's code is now written by AI. But Xu doesn't think that's enough, because writing code is only a slice of what engineers actually do.

Flux effectively stretches that AI automation from "write the code" to the entire engineering workflow, letting agents handle code review, CI troubleshooting, and maintenance in the background.