DoorDash’s Flux platform ran 130,000 engineering tasks in a single month, moving background agents off developers’ laptops and into managed cloud environments. The system also supports more than 25,000 automated code reviews each week, with over 300 reusable playbooks and more than 10,000 weekly workflow invocations.
The company built Flux because laptop-based agents are constrained by local compute, disappear when a device disconnects and can inherit broad access to a developer’s credentials. Cloud execution gives DoorDash a central place to define permissions, observe activity and stop work that crosses policy boundaries.
Each task runs in an isolated Firecracker micro virtual machine provisioned with the required repositories, tools, secrets and dependencies. DoorDash reports a 95th-percentile setup time below five seconds. An internal Model Context Protocol gateway controls access to company systems and logs actions, while YAML playbooks specify tools, validation steps, permissions and safety limits.
Engineers can start workflows from Slack, GitHub, scheduled jobs, a command line or conversational interfaces. After beginning with code review, DoorDash expanded Flux to continuous-integration triage, on-call work, maintenance and ticket-driven development. The scale is notable, but the architecture’s bigger lesson is that agent adoption quickly becomes an identity, isolation and audit problem rather than only a model-selection problem.