AI Engineer: "All Code Is Slop (Except Mine) — Lauren Tan, SpaceXSI". Facts on this page cite the primary sources below, not the video.

The claim: AI agents now write most of Grok Bot's code, and the quality went up, not down.

The proof: On Wednesday, Nov. 11, 2026, Lauren Tan, a principal engineer at SpaceX SI who works on Grok Bot, gave a talk called "All Code Is Slop (Except Mine)" at the AI Engineer Code Summit in San Francisco. AI Engineer posted the 17-minute recording on Saturday, Nov. 14. Her numbers: 11,000 pull requests shipped to production last month, a new stable app version about once a day, 227 people contributing in October, and a revert rate steady at about half a percent. Those are her team's figures. Nobody outside has checked them.

Editorial Thesis

Most people try to control AI by writing longer instructions and hoping. Tan's talk is the opposite pitch: instructions get ignored, rules don't. Every time her team corrects an agent, they try to turn that correction into something the code enforces, a lint rule, a test or a better design, so the same mistake can't land again. That is a management idea as much as a coding one, and it's the most copyable part of the talk. The 11,000 is the headline. The roughly 6 months of setup it took her is the actual job.

Honest Verdict

  • Confirmed (AI Engineer's own recording and talk page): the talk happened on Nov. 11 and the video went up on Nov. 14. The 11,000 PRs, about-once-a-day releases, 227 contributors in October and ~0.5% revert rate are all stated in it.
  • Confirmed (live transcript): "last month, our team shipped a ridiculous 11,000 pull requests to production" and "most product code that we land no longer really requires deep human review."
  • Her description, not an audit: AI Engineer's write-up and Sulat both note these are the team's reported outcomes, not a controlled test.
  • Not defined in the talk: what counts as a revert, over what window, or how many of the 11,000 were written by agents versus people.
  • Not shown: a revert rate from before the team leaned on agents, so "it got better" can't be measured from outside.

The numbers in human terms

Her numberWhat it meansSource
11,000 pull requests last monthAbout 1 every 4 minutes, day and night, across October's 31 days (our math)Talk, transcript
~0.5% revertedAbout 1 in 200 changes got rolled backTalk
227 contributors in OctoberIncluding drive-by contributors with little contextTalk, transcript
Stable release about once a dayA new desktop app build most daysTalk, transcript
About 6 monthsHow long she says it took to reach thousands of PRs, including while asleepAI Engineer talk page

How they say they do it

Rules over reminders. Tan draws an upside-down pyramid. Most of the effort goes into the code itself: good APIs and architecture. Next come lint rules, types, compilers and tests. Then verification. Written guidance, like an AGENTS.md file, gets "a little bit." Her rule of thumb: when you correct an agent, make the correction a deterministic check, not one more line of prose.

Banned on purpose. The team's in-house framework, Dune, bans React's useEffect in product code and blocks imports between the Electron main and renderer folders. It also bans about 99% of code comments. Tan's reason: agents wrote comments to justify quick timer hacks, and other agents copied the hack along with the comment. Now the lint error tells the agent to fix the root cause.

Review the evidence, not the diff. Code review on her team has shrunk to checking videos, screenshots, diagrams and formal checks. They model user flows in TLA+, prove properties of a library in Lean, and run Bombadil to poke at the interface like a user would. Tan says most of her own AI token use now goes to verification.

Then automate. Only once those foundations hold, she says, does it make sense to let agents run loops on their own. Her example: when the team cuts a release, Grok Bot starts a Cursor project that fuzzes the app, opens a fix for each bug it finds and merges it automatically.

What this does not prove

  • It does not prove AI wrote the 11,000. Tan didn't break the number down by author or size.
  • It does not prove the method works for everyone. She says it took her about 6 months, and her team built its own framework.
  • It does not prove Grok Bot is bug-free. A 0.5% revert rate still means dozens of rollbacks in a month at that volume.
  • It isn't independent. Every figure here is the Grok Bot team describing itself.

What to watch next

  1. Whether SpaceX SI publishes a definition of "revert" or the breakdown of agent-written versus human-written changes.
  2. Whether the Dune framework or its lint rules get released (only her pstack plugin is public).
  3. How the Grok Bot changelog pace holds up after 1.0.

Bottom Line

Lauren Tan says the Grok Bot team shipped 11,000 pull requests in October with a revert rate around 0.5%, and that the trick is turning every correction into a rule the code enforces. It's a strong, specific talk and a useful playbook. The numbers are the team's own, so treat them as a claim with a recording behind it, not an audit.

Source notes

Talk recording and description from AI Engineer on YouTube (posted Nov. 14). Quotes checked against the conference's live transcript of the Nov. 11 session. Framework details (useEffect, comments, Electron folders, TLA+, Lean, Bombadil) from Sulat's Nov. 16 write-up of the talk. The "1 every 4 minutes" and "1 in 200" conversions are our arithmetic from her figures. Video: AI Engineer (official conference channel); alternate: GrokBot Guru explainer (Nov. 16).

Sources