AI Pulse · Under the Radar

DoorDash is running 130,000 automated engineering tasks every month through its own agent platform. That’s not a demo or a pilot. It’s production work at a company with real security constraints and compliance requirements.

The breakdown matters. They’re running more than 25,000 automated code reviews each week. They’ve built over 300 reusable “playbooks” that get invoked more than 10,000 times a week. Every agent runs in a sandbox with scoped access, and every action is audited. This is what an internal cloud-agent platform looks like when you can’t afford to let an AI roam freely through your codebase.

What makes this useful as a precedent is that DoorDash isn’t selling anything. They built this for themselves, which means the architecture reflects actual constraints, not marketing claims. The playbooks are templates for common engineering tasks. Code reviews, testing workflows, deployment checks. The kind of repetitive work that eats up senior engineer time but still needs human oversight before it ships.

The security model is straightforward. Each agent gets limited permissions tied to the task it’s running. Actions are logged. If something breaks or behaves unexpectedly, you can trace it. That’s the kind of setup you need if you’re running agents in an environment where a mistake could take down a delivery platform or expose customer data.

This is the sort of thing we build into an AI command centre. Not the agents themselves, but the orchestration layer that lets you run them safely, track what they’re doing, and pull the plug if needed. DoorDash had the engineering resources to build their own. Most companies don’t, but the pattern is now clear. Agents at scale need guardrails, audit trails, and a way to turn repetitive tasks into reusable workflows. The technology works. The question is whether your infrastructure can support it.