DoorDash Flux Handles 130,000 Tasks/Month via Cloud Agent Platform
In this article
DoorDash has shifted AI coding agent workloads off developer laptops onto a centralised cloud platform called Flux, which processed 130,000 engineering tasks in a single month in 2026. The platform handles more than 25,000 automated code reviews per week, with over 300 playbooks and more than 10,000 weekly playbook invocations — figures that make the operational case for cloud-native agent infrastructure far more concretely than most published accounts. For engineers thinking about pipeline architecture rather than better models as the driver of AI productivity gains, Flux is a live data point worth dissecting.
Why Laptop-Based Agents Hit a Ceiling
DoorDash's original agent setup ran on individual developer machines, imposing hard constraints on three dimensions simultaneously: local CPU and memory bounded workload size; execution depended on the developer's device staying connected; and autonomous agents inherited that developer's full credential scope, creating an audit surface that was difficult to scope or observe. DoorDash cited the inability to monitor which systems agents accessed, under whose authority, and across which concurrent sessions as a core driver of the migration. Security architect Radoslav Krehlik noted that the move to cloud brought "full enterprise security guardrails" that local execution structurally cannot provide.
The Four Platform Primitives
Flux is organised around four components: cloud sandboxes, an MCP gateway, reusable playbooks, and invocation surfaces.
| Primitive | Technology / Implementation | Primary Role |
|---|---|---|
| Cloud Sandboxes | Firecracker micro-VMs | Isolated execution; repos, secrets, and build tools provisioned per task |
| Agent Gateway (MCP) | In-house MCP gateway | Scoped permissions to internal systems; activity logging for auditing and policy enforcement |
| Playbooks | YAML definitions | Specify task, required tools, permissions, validation, and safety boundaries; mix agent-driven and deterministic steps |
| Invocation Surfaces | Slack, GitHub, cron, CLI, conversational interfaces | Allow workflows to be triggered unattended and in parallel without developer intervention |
The Firecracker micro-VM choice is significant: DoorDash reports a 95th-percentile service level objective of under five seconds for end-to-end sandbox setup, covering microVM start, repository clone, build tool installation, and coding agent harness configuration. That sub-five-second p95 target makes parallelism at tens of thousands of weekly invocations practical rather than theoretical. The Agent Gateway enforces scoped permissions rather than inheriting ambient developer credentials and generates audit logs that make agent activity reconstructable after the fact.
Workflow Expansion and Observability
DoorDash began with automated code review as the initial Flux use case — frequent and measurable enough to validate the platform before expanding scope. The team subsequently extended Flux to CI triage, on-call workflows, maintenance tasks, and ticket-driven development. This staged expansion mirrors the pattern of using a high-volume, low-ambiguity workflow to stress-test infrastructure before routing more complex work through it.
One architectural decision worth noting is the shift from private Slack channels to public threads for agent execution notifications. The change was intentional: public threads allow engineers to observe agent runs, review outputs, and see how other teams are delegating work — an observability and organisational learning mechanism that surfaces agent behaviour to the broader engineering population rather than keeping it siloed per team.
Duy Nguyễn, Co-founder of TOP GROUP Vietnam, framed the resulting challenge clearly: once agents handle background work, code review, and concurrent execution at scale, "the hard problem shifts from model choice to control: identity, permissions, sandboxing, audit logs, quotas, and a real kill switch." The credential-exposure risks that arise in contexts like AI coding agent credential hijacking are precisely what Flux's gateway layer is designed to structurally prevent.
The Broader Infrastructure Signal
GitHub supports both local and cloud sandbox environments for Copilot, with controls over filesystem, network, and system access — indicating that cloud-based agent execution is becoming a recognised infrastructure category rather than a DoorDash-specific experiment. What makes Flux notable is the operational volume behind the design choices: 130,000 tasks in a month, 25,000 reviews per week, and more than 10,000 playbook invocations weekly are not pilot numbers. They represent a production-grade argument that the architectural layer separating agent invocation from agent execution is the leverage point determining whether agentic workflows scale safely — not the underlying model. Organisations evaluating similar infrastructure should treat the p95 sandbox latency target, the MCP gateway's permission scoping, and the YAML playbook schema as the decisions that set ceiling throughput, not the choice of coding agent itself.