DoorDash Transitions Engineering Work to Cloud-Based AI

Food delivery giant DoorDash has migrated its internal engineering work from developers' personal PCs to the cloud platform 'Flux'. With this system in full operation, 130,000 engineering tasks have been automatically processed in one month, and over 25,000 code reviews are being automated weekly.

Why was it necessary to move AI processing from developers' PCs to the cloud in the first place? AI agents—autonomous AI programs that make decisions and execute tasks—perform complex consecutive operations such as writing code and running tests. Executing these on individual developers' PCs can create instability due to machine specifications and environmental differences, and it becomes difficult to maintain records of who performed which processes. By consolidating to the cloud, these issues can be comprehensively resolved.

Flux's architecture employs several distinctive technologies. First, it uses a lightweight virtual environment called 'Firecracker microVM' to create independent execution spaces for each task, thereby minimizing security risks. Firecracker was originally developed by cloud providers for large-scale server operations and is known for fast startup and minimal resource waste. It also manages integration with external tools and services through an 'MCP Gateway' and standardizes agent behavior by combining reusable 'playbooks' (operation manuals). Moreover, by possessing multiple invocation pathways, it is designed to enable agent activation from various scenarios.

Security and audit considerations are also evident in the design. It adopts 'scoped access' that restricts each agent's accessible range per task and incorporates a 'centralized auditing' system that manages all processing records in one place. AI agents that access corporate internal systems always face risks of operational errors and misuse. This design aligns with the important principle of minimizing privileges granted to AI while enabling traceback of what occurred—a crucial approach for enterprises to safely operate AI.

The significance of this development extends beyond DoorDash alone. As a case study of transitioning from an experimental stage where AI agents are used as 'developer tools' on personal PCs to a stage where they are integrated into production environments as enterprise infrastructure, it can serve as a reference for the industry as a whole. The scale of 130,000 tasks monthly and over 25,000 weekly can be seen as a figure demonstrating that AI is beginning to assume part of routine engineering work beyond its supplementary role.

A key point to watch going forward is the question of how far this infrastructure can be scaled. As the number of tasks and scope of automation expand, the impact of errors in agent judgment increases accordingly. How DoorDash develops its governance and oversight framework for Flux will be a perspective relevant not only to DoorDash but also to other enterprises seeking to build similar platforms.