Eleven thousand pull requests in one month sounds less like a software team and more like a denial-of-service attack on code review. Yet that is the number Lauren Tan, a principal engineer on Grok Bot at SpaceX’s AI unit (branded SpaceXSI on its docs, SpaceXAI in many posts), put on stage at the AI Engineer Code Summit. Her talk, “All Code Is Slop (Except Mine)”, went online on November 14, the same day Y Combinator published a long interview with Tan and Roman Ugarte about building Grok Bot from the first line of code to launch in seven weeks. Together they offer a rare look at how an AI agent product is built largely by AI agents, and what keeps that from turning into chaos.
Quick Take
- The claim: The Grok Bot team shipped 11,000 pull requests (proposed code changes) to production in October 2026, with about 227 people contributing at least one, and a new stable release roughly once a day, Tan said.
- The quality figure: About 0.5% of pull requests get reverted, a share Tan says has fallen as volume rose. These are the team’s own numbers and have not been independently verified.
- The method: Every correction to an agent becomes a lint rule or an architecture change instead of a note in an instructions file. Review shifts to “verification artifacts” such as videos, screenshots and formal proofs.
- The product story: Grok Bot began as an internal prototype built in about seven weeks. It gives AI “teammates” their own identities, logins and computers and keeps them running in the background.
- Why it matters: It is a detailed, public recipe for running coding agents at scale, from a team that sells an agent product, so it is both a playbook and a pitch.
In plain English
Grok Bot is a product from SpaceX’s AI division that lets you create AI helpers that work like digital colleagues. One of its engineers explained in a talk how her team now lets AI write most of the product’s code, about 11,000 changes in a single month. The trick, she says, is not to trust the AI blindly, but to build automatic checks into the code itself so that bad work gets caught before it ships. The team also makes the AI prove its work with screen recordings, screenshots and mathematical checks. Only about one change in two hundred later has to be undone, according to the team.
What happened
Tan gave the talk on the main stage of the AI Engineer Code Summit on November 11 and posted the recording on November 14: “my AIE talk is up! come hear how the grok @bot team ships so quickly without compromise on quality.” The session description sums up the premise: “The Grok Bot team at SpaceXAI lands hundreds of PRs a day. Nobody reads them all, yet quality goes up anyway.”
From the talk transcript: “last month, our team shipped a ridiculous 11,000 pull requests to production, which is a wild number.” Tan added that “in October, we had about two twenty seven people who contributed at least one pull request” and that the number of reverted pull requests “has actually gotten lower” in relative terms, “steady at about half a percent.” A slide on performance showed the app’s cold launch time falling from 2,304 ms in May to 1,395 ms (minus 39%), even though, as Tan noted, “Grokbot is still an Electron app for now.”
The same day, Y Combinator released a Lightcone episode with Tan and Ugarte, who “are on the core team that took Grok Bot from first line of code to launch in seven weeks.”
How the team does it
Tan’s recipe, as laid out in the talk:
- Turn corrections into code. “Every time we correct our agents, it turns into one of these deterministic checks,” Tan said, or the team removes the problem “through better API design, through better architecture.” Lint rules in the team’s own Electron framework, called Dune, return error messages that tell the agent “to fix the root cause.” According to an AI Aulis summary of the talk, the team also banned 99% of code comments, because agents used them to justify quick fixes.
- Verify, then review the evidence. Code review “has now really shrunk into looking at verification artifacts instead. So things like videos, screenshots, diagrams, formal verification.” Agents run the app themselves with debugging tools such as the Chrome DevTools Protocol, guided by a “feature map” of the app from the user’s view.
- Add formal methods where it counts. The team uses TLA+ to model-check flows such as notifications across devices and Lean to prove properties of its own virtualization library. “We’ve already found quite a few bugs,” Tan said, while calling it “not a silver bullet.” A UI property-testing library called Bombadil caught a few more.
- Close the outer loop. After each release, Tan said, Grok Bot starts a Cursor project that fuzzes the whole app like a real user, opens fix pull requests for bugs it finds and auto-merges them.
Her warning for everyone else: “if your inner coding loop produces slop, then automating that loop just produces more slop.” Tan’s skills package, pstack, is published in Cursor’s plugins repository on GitHub.
Background: seven weeks to an AI teammate
In the YC interview, Ugarte described the starting point: “we wanted to build a product where the work was organized around bots as the key operative principle, not around chats or around files.” The bots should have “their own computer,” use the same tools a human does and “always be on.” Internally, he said, the marketing team gave a bot its own login to ad platforms and a budget, and the data science team built a read-only “data bot” in Slack to answer analysis requests.
Tan described building a feature end to end in a Slack thread of about 200 messages with a colleague and her bot, without opening the code editor. Ugarte said Tan is now “putting up 4,000 PRs a month,” and that the new bottleneck is “alignment among us as a team of like what do we want the product roadmap to be.” His rule for that: “if something doesn’t get internal product market fit, we don’t ship it.”
i10x has followed Grok Bot’s rise, from agent teams with a “chief of staff” bot to the token limit and governance debate. This week the product also gained the ability to join video meetings on Google Meet, Zoom and Microsoft Teams.
Pros and cons
For users
- Pro: A team shipping daily releases with automated bug fixing can respond to bug reports fast. Tan says the app got faster, not slower.
- Con: Ugarte himself warned that frequent changes to “the core pixels” frustrate users, which is why the team says it holds back features that do not win internally first.
For developers
- Pro: A concrete, free playbook: lint against every recurring agent mistake, give agents a way to run the code, review artifacts instead of diffs, and add formal checks where bugs hurt most.
- Con: It took time. The AI Aulis summary notes it took Tan six months to reach thousands of production pull requests, and the setup assumes deep investment in tooling plus, as Tan said in the YC interview, “unlimited tokens” at a lab.
For companies
- Pro: Shows agents can raise output without raising revert rates, if the codebase enforces quality, and that non-engineering teams can own bots for their own functions.
- Con: Pull request counts are a weak productivity metric, and self-reported. Letting agents merge their own code is a governance decision most regulated companies are not ready to make.
What X is saying
Tan’s own post kicked things off:
my AIE talk is up! come hear how the grok @bot team ships so quickly without compromise on quality
— lauren (@poteto) November 14, 2026 · View on X
Y Combinator framed the product side:
At @spacexai, Lauren Tan (@poteto) and Roman Ugarte (@romanugarte_) are on the core team that took Grok Bot from first line of code to launch in seven weeks.
— Y Combinator (@ycombinator) November 14, 2026 · View on X
It lets you spin up AI teammates with their own identities and logins, and they just keep working in the background.
Attendee @rae1101x summarized two more numbers from the talk: Tan “went from 68 PRs in march to 3,127 in october,” and the team’s “revert rate went down, not up (0.98% → 0.47%).” Not everyone was impressed by volume as a metric. “11k PRs sounds terrible,” replied @0x_fokki. Quality engineer @OffCamberQE asked the question many engineering managers will: “Curious how you’re measuring quality beyond reverts and performance. Are you tracking escaped defects or change failure rate?”
Users, meanwhile, praised the product the code ships into. “Grok Bot has been such a force multiplier for me. It’s insane,” wrote Adeo Ressi.
Grok’s take
Asked about the video, Grok summarized (translated from Chinese) that Tan explains how to keep agents productive without lowering quality by “turning every mistake into lint or architecture constraints rather than prompts,” with codebase design, static analysis and verification, including TLA+ and Lean, as the core.
Critical analysis
The talk’s strongest idea is also its least flashy: instructions files rot, code does not. A lint rule that fires exactly when an agent repeats a mistake is cheaper and more reliable than a growing list of “please don’t” in a prompt. That advice is portable to any team using coding agents, regardless of vendor.
The numbers deserve more caution. 11,000 pull requests say little without knowing their size, and a 0.5% revert rate measures only what was noticed and rolled back, not bugs that slipped into production quietly, which is exactly what @OffCamberQE asked about. The figures come from the team itself, in a talk that ends with a pitch for Grok Bot as “a great way to build that outer loop.” Tan is also candid that her setup relies on lab-level token budgets. Smaller teams should copy the method, not the volume.
There is one more claim floating around the trend: that SpaceX’s AI unit is shifting focus away from image tools such as Grok Imagine toward agents. i10x found posts from users noticing the Imagine tab disappearing, and others saying it is being folded into the main Grok chat, but no official statement. Treat it as unconfirmed.
What’s still open
- How the team measures escaped defects and change failure rate beyond reverts.
- How large the average pull request is, and how many are agent-authored versus human-authored.
- Whether Grok Bot will offer the release-fuzzing “outer loop” Tan described as a packaged feature for customers.
- What SpaceX’s AI unit plans for Grok Imagine. There is no official word.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis is based on the recording and transcript of Lauren Tan’s AI Engineer Code Summit talk, Y Combinator’s Lightcone interview with Tan and Roman Ugarte, an AI Aulis summary of the talk, SpaceXAI’s Grok Bot documentation and public posts on X. All productivity figures are the team’s own and were not independently verified.
i10x Perspective
“All code is slop (except mine)” is a joke with a serious point: trust in AI-written code does not come from better prompts but from a codebase that refuses bad code by design. The Grok Bot team is effectively selling the outer loop and showing off the inner one. Whether 11,000 pull requests is impressive or alarming depends on metrics nobody outside the team can see yet. But the method (lint every mistake, demand proof, automate only what is already good) is the most useful thing a lab engineer has said about coding agents this month, and it costs nothing to try.






