DoorDash's Flux Runs 130,000 Engineering Tasks Through Cloud-Based Agents

DoorDash migrated engineering agent workloads from developer laptops to its Flux cloud platform because local execution has multiple limitations. Local execution is constrained by available CPU and memory, depends on the developer's device staying connected, and may allow autonomous agents to access the developer's existing credentials and internal systems. Additionally, local execution makes it harder to monitor where agents run, what systems they access, and on whose behalf they operate.

DoorDash security architect Radoslav Krehlik stated on LinkedIn that Flux was built after laptop-based agents hit power, uptime, and security limitations. Engineers can trigger agents from Slack, GitHub, or scheduled tasks, while DoorDash maintains full enterprise security guardrails.

Flux is built around four platform primitives: cloud sandboxes, an MCP gateway, reusable playbooks, and invocation surfaces. Playbooks define the work, sandboxes provide isolated execution environments, the gateway controls access to internal systems, and invocation surfaces allow workflows to be started via Slack, GitHub, cron, command line, or conversational interfaces.

Sandboxes use Firecracker microVMs to isolate agent workloads. Each environment is pre-provisioned with the repositories, development tools, secrets, and runtime dependencies needed for the task. DoorDash reports 95th percentile end-to-end sandbox setup time under 5 seconds, including starting the microVM, cloning the repository, installing build tools, and configuring the coding agent framework.

Agents access DoorDash internal systems through the Agent Gateway (an internal MCP gateway), which provides scoped permissions and logs agent activity for audit and policy enforcement. Playbooks are defined in YAML, specifying tasks, required tools, permissions, validation, and security boundaries, and can combine agent-driven steps with deterministic code.

Flux was initially used for automated code review, which DoorDash describes as a frequent and measurable workflow. It then expanded to CI triage, on-call tasks, maintenance workflows, and ticket-driven development. The platform supports over 300 playbooks, with over 10,000 invocations per week, and workflows can run unattended in parallel.

DoorDash also changed its Slack integration from private channels to public threads, allowing engineers to observe agent execution, review results, and learn how other teams delegate work. This transparency helps build trust and collaboration.

Duy Nguyễn, co-founder of TOP GROUP Vietnam, commented on LinkedIn that once agents handle background work, code review, internal tools, and concurrent execution, the hard problem shifts from model selection to control: identity, permissions, sandboxing, audit logs, quotas, and a real kill switch.

This view aligns with Flux's design: scoped permissions and audit through the MCP gateway, isolation through Firecracker sandboxes, and security boundaries defined by playbooks. Control mechanisms become key to large-scale agent deployment.

Flux's approach is part of a broader movement of coding agents moving to cloud execution. GitHub supports both local and cloud sandbox environments for Copilot, with controls over file system, network, and system access. Its cloud sandbox provides an isolated environment for agents to perform coding tasks without consuming local developer resources.

This suggests that as agents take on more critical tasks, cloud execution and centralized control may become standard practice, driven by the security and resource limitations of local execution.