SpaceX Super Intelligence's Grok Bot engineering team uses AI coding agents in its software development workflow. In a November 2026 AI Engineer talk, engineer Lauren Tan reported that the team shipped 11,000 pull requests to production in the preceding month, with a revert rate of about 0.5%. These are team-wide figures; she did not say all pull requests were generated by AI.
Tan described spending six months building an environment in which agents could work with greater autonomy while the team maintained code quality. The workflow combines code architecture, static checks, tests and runtime verification. Engineers review verification artifacts such as screenshots, videos and diagrams rather than relying solely on prose explanations from agents.
For the Grok Bot desktop client, the team developed a framework called Dune with explicit architectural constraints and lint rules. Product code restricts use of React's useEffect; separate Electron main and renderer directories are checked to prevent prohibited imports. The team also uses lint feedback to direct agents toward fixing root causes rather than adding comments to justify shortcuts.
For runtime verification, agents can run the application and use developer tools to inspect behavior. The team has also experimented with TLA+ model checking for user flows, Lean proofs for algorithms, and property-based UI testing. Tan said formal verification had already uncovered bugs, while describing it as one layer rather than a complete solution.
After a release, Grok Bot can start a Cursor project to fuzz the application, identify bugs and open fix pull requests; Tan said the workflow can automatically merge those fixes. She reported that the team released a stable client roughly daily and that the application's startup and interaction performance had improved. The PR and revert figures are self-reported team outcomes, not an isolated measurement of the agents' impact.
Details
- City
- San Francisco
- Organization
- SpaceX
- Continent
- North America
- Country