Grok Bot Crew Clears Eleven Thousand Agent-Written PRs in One Month
Voicing this story…
SpaceX Super Intelligence just showed what “coding agents at scale” actually looks like. On November 14 the talk “All Code Is Slop (Except Mine)” went live. Lauren Tan, principal engineer on the Grok Bot team, walked through numbers that stopped the AI engineering crowd cold.
In one month—October—the team merged 11,000 pull requests into production. Roughly 0.5 percent got reverted. About 227 people landed at least one change. Tan herself accounted for around 3,000. The client ships a new stable build almost every day.
Those are the team’s own figures from the AI Engineer Code Summit stage. No public repo lets outsiders audit them, yet the method is more interesting than the raw count.
CONSTRAINTS BEAT PROMPTS
Tan’s core rule is simple: every time an agent messes up, turn the fix into an enforceable check inside the codebase instead of another line in a rules file. Agents ignore documentation. They cannot ignore a lint rule that fails the build.
The team’s Electron framework, Dune, bans React’s useEffect in product code because agents misuse it. Main-process and renderer code live in separate folders with a hard dependency block after agents kept creating jank. Ninety-nine percent of comments are outlawed; agents used them to paper over shortcuts.
Code review has shrunk to checking verification artifacts—screen recordings, screenshots, diagrams. Formal tools fill the gaps: TLA+ models multi-device notification flows, Lean proves properties of their virtualization layer, and a property-based UI tester called Bombadil catches edge cases. Tan says most of her token budget now goes to verification, not generation.
AGENTS THAT FIX THEMSELVES
After each release Grok Bot spins up its own Cursor project, fuzzes the app like a real user, opens fix PRs, and auto-merges many of them. The outer loop is closed. Cold-start time dropped from 2,304 ms in May to 1,395 ms even while the app is still Electron.
Grok Bot itself started as an internal prototype. Tan and colleague Roman Ugarte took it from first line to launch in seven weeks. Bots get their own identities, logins, and always-on computers. Marketing gave one an ad budget. Data science built a read-only Slack bot. Tan once finished a full feature inside a 200-message Slack thread with a bot and never opened the editor.
The same pattern now powers the product customers use. The bottleneck has moved from writing code to deciding what is worth building. Features that lack internal product-market fit simply do not ship.
This lands the same weekend Grok Imagine folded into the main chat after reports of heavy staff cuts on the image side. The coding-tools-and-agents shift is no longer a rumor. It is 11,000 merged changes, a half-percent revert rate, and a public playbook that other teams are already copying. The sky, as Musk keeps saying, is not even close to the limit.


0 COMMENTS