Grok · Grok Bot · 15 Nov 2026
The Grok Bot team says it shipped 11,000 pull requests in a month. Here's how, and what the numbers leave out
Lauren Tan, an engineer on Grok Bot at SpaceXSI, says her team merged 11,000 pull requests to production in a single month, that about half a percent of them were reverted, and that the app got faster while it happened. She shipped 3,000 of them herself. It took six months of work on the setup, she said, before she could ship thousands, "a lot of them in my sleep." Her talk, posted on Saturday, explains the setup. These are the team's own numbers from a conference stage, and nobody outside the company can check them.
On this page
The editor holds Tesla and SpaceX shares and uses Grok Bot. Supercharged Daily is independent and not affiliated with either company. Nothing here is financial advice.
Video: "All Code Is Slop (Except Mine)," Lauren Tan, SpaceXSI. AI Engineer on YouTube (17 minutes 31 seconds). Embedded from YouTube. Timestamps below refer to this video.
What she said
Tan posted the talk at 1:36 p.m. ET on 14 Nov (10:36 a.m. PT), writing that it shows "how the grok @bot team ships so quickly without compromise on quality." The video runs 17 minutes 31 seconds. The headline numbers come near the end, at 13:37: "last month, our team shipped a ridiculous 11,000 pull requests to production." She added that the team now cuts a new stable version of the client "about once a day," and that in October about 227 people contributed at least one pull request, many of them drive-by contributors outside the core team.
On quality, she said reverted pull requests have fallen on a relative basis and are "steady at about half a percent" (14:21). She also said startup, typing, opening new chats and switching chats have all gotten faster since launch, even though Grok Bot is still an Electron app "for now."
The numbers, worked through
| Figure from the talk | What it works out to |
|---|---|
| 11,000 pull requests in a month | About 355 to 367 merged changes a day, depending on whether the month had 30 or 31 days |
| About 0.5% reverted | Roughly 55 pull requests rolled back out of 11,000 |
| About 227 contributors in October | About 48 pull requests per contributor, on average, if the 11,000 month was October |
| Tan's own 3,000 in a month | About 27% of the team total from one engineer |
| Code review in 2024 vs now | In 2024 she'd go into work and review the three pull requests her colleagues wrote. "Today, that number is more like 300" |
She said it took her six months of work on the setup before she could ship at that rate. The per-day, per-contributor and share figures are my math.
How they say they do it
Tan described an inverted pyramid of where to spend effort. The wide top is the code itself: API design, architecture, lint rules, type systems and tests, which stop an agent from doing the wrong thing in the first place. Below that sits verification, at runtime and formally. Written guidance like rules files, skills and agents.md comes last, because it needs judgment and an agent can ignore it.
Her main rule is that every time the team corrects an agent, the correction becomes a check the code enforces, instead of another line in a rules file. Some examples from the talk:
- Grok Bot runs on a framework the team built called Dune, which Tan compared to "Next.js for Electron apps." React's
useEffect, which she said agents misuse, is banned in product code (5:52). - Code for the app's main process and its window renderer lives in two folders, and a dependency checker blocks imports between them, after agents working on Cursor, another Electron app, kept importing main-process code into the renderer and causing jank (6:12).
- The team banned "99% of comments" (8:20). Tan said agents used comments to justify quick fixes, and a comment can't be enforced.
- Code review has mostly turned into checking evidence: videos, screenshots, diagrams. The team uses TLA+ to model-check notifications across devices, Lean to prove properties of its custom virtualization library (flicker and jitter), and a property-based UI tester called Bombadil (10:52). She said they are still experimenting with the formal-verification work.
She said most of her own token usage now goes to verification. At 16:20 she described the end state: on each new release, Grok Bot starts a Cursor project that uses the app like a real user, finds bugs, opens fix pull requests and merges them automatically. Most product code the team lands "no longer really requires deep human review," she said.
What the numbers don't tell you
- All of this is self-reported in a conference talk. There's no public dashboard or repository to check it against.
- The talk doesn't date "last month." The 227-contributor figure is explicitly October, so I've only paired the two where noted.
- A revert rate counts changes that were pulled back. A bug that ships and gets fixed by a new pull request instead wouldn't show up in it.
- Pull request counts measure volume. Agents tend to make many small changes, so 11,000 small PRs and 11,000 large ones are very different amounts of work.
- The speed claims (startup, typing, chat switching) came without figures in the talk.
Why it matters if you use Grok Bot
If the numbers hold, a new client every day or so explains why features like team bots in 1.0 keep landing quickly, and why users see small changes week to week. The routine she described, where Grok Bot launches coding agents to test each release, is the same routines-plus-cloud-agents setup subscribers can build. Her advice for anyone copying it: fix the codebase first. "If your inner coding loop produces slop, then automating that loop just produces more slop," she said, and you end up throwing more tokens at it.
my AIE talk is up! come hear how the grok @bot team ships so quickly without compromise on quality https://t.co/Fel7tlxqSL
— lauren (@poteto) November 14, 2026
Would you trust an app where agents merge most of the code, if the revert rate really is half a percent?
— Alex R
Sources: "All Code Is Slop (Except Mine)," Lauren Tan, SpaceXSI, AI Engineer on YouTube (quotes and timestamps from the talk; video is 17 minutes 31 seconds, uploaded 14 Nov 2026 at 1:30 p.m. ET); @poteto on X, 14 Nov 2026. Per-day, per-contributor and share calculations are my own. Video: AI Engineer, embedded from YouTube. Embed: Lauren Tan (@poteto) on X. Card photo: still from "DotJS 2018 — Lauren Tan — Learning to Love Type Systems", CC BY 3.0 (dotconferences), filmed 9 Nov 2018 in Paris; this is not a still from the 2026 talk, and the conference lower-third has been cropped off.
