Agentic AI
DoorDash Moves AI Agents to Cloud, Automating 130,000 Tasks Monthly
The delivery giant's Flux platform orchestrates engineering work across isolated cloud environments instead of developer machines, handling 25,000 code reviews weekly with enterprise-grade security controls.
InfoQ
DoorDash has transitioned autonomous engineering agents from individual developer laptops to a centralized cloud platform called Flux, which executed 130,000 engineering tasks in a single month during 2026. The shift addresses the resource constraints, security risks, and observability challenges that come with running AI workloads on local machines.
Why Cloud Over Laptops#
When agents operate on personal devices, they're constrained by the machine's CPU and memory, dependent on the developer keeping their computer online, and often inherit broad access to internal credentials and systems. DoorDash found that tracking where agents execute, which systems they touch, and whose behalf they act on became increasingly difficult. The company determined that a dedicated cloud infrastructure could solve these problems while enabling unattended, parallel execution at scale.
How Flux Works#
The platform rests on four core building blocks: isolated cloud sandboxes, an access control gateway, reusable workflow templates, and multiple trigger points. Playbooks written in YAML describe the specific work to be performed, including required tools, permissions, and safety constraints. Sandboxes use Firecracker micro virtual machines to create isolated execution environments, each pre-configured with repositories, development tools, secrets, and runtime dependencies. DoorDash achieves a 95th percentile setup time of under five seconds—from microVM startup through repository cloning to agent initialization.
Agents communicate with internal systems through a custom MCP gateway that enforces scoped permissions and logs all activity for audit trails and policy compliance. Engineers trigger workflows via Slack, GitHub, cron jobs, the command line, or conversational interfaces.
Growing Use Cases#
DoorDash began with automated code review, a frequent and quantifiable workflow, then expanded Flux to handle CI triage, on-call tasks, routine maintenance, and ticket-driven development. The platform now powers more than 300 playbooks and sustains over 10,000 weekly invocations, including 25,000 automated code reviews per week. The team shifted its Slack integration from private channels to public threads, enabling engineers to observe agent executions, learn from peer workflows, and identify new automation opportunities.
