They even gave a bot its own ad budget: the Grok Bot team that shipped 11,000 PRs in a month says "fix agent mistakes in code, not prompts"
In 3 lines
- Bots get names, cards and computers, and work moves from task to outcome to job; marketing even handed one an ad budget
- The team shipped 11,000 PRs in a month with ~0.5% reverted, by turning every agent mistake into a lint rule
- With execution abundant, the bottleneck moved to deciding what and why, while permissions, cost and revenue remain open

Y Combinator's Lightcone podcast hosted Lauren Tan and Roman Ugarte, the people behind Grok Bot, in an episode released on November 15. Grok Bot is the 'AI teammate' product from SpaceXSI, the renamed xAI, and each bot gets its own name, phone number, payment card and computer. The pair took it from first line of code to internal launch in seven weeks, and the marketing team now hands bots ad accounts and a budget. We read the 60-minute conversation, which passed 130,000 views in three days, alongside Tan's talk that hit 810,000 views the same week, launch coverage and critical takes.
An 'AI teammate' built in seven weeks
Ugarte said the team saw how AI coding had completely transformed the way its engineers worked and asked what that would look like for non-engineers. They landed on three principles: organize work around bots rather than chats or files, give bots the same tools humans use and their own computer, and keep them always on so they come to you proactively. When it was released internally, both the most AI-forward and some of the least AI-forward people in the company enjoyed it. Grok Bot went through a beta starting in August and reached version 1.0 on November 11.
From a task to an outcome to a whole job
Usage tends to widen in the same order, he said: people try a small task, then give an outcome or objective, then hand over a full job. The company's performance marketing team gave its bot logins to Google Ads and Facebook Ads plus a budget, and had it A/B test assets made by the design team's bot. Surprisingly, some of the most enthusiastic users are small businesses. Companies that struggle to hire can have a new bookkeeping 'employee' within hours.
Many bots that each care about one thing
Ugarte compared setting up bots to hiring. Rather than one superhuman who cares about a hundred things at once, you want someone obsessed with one problem, even in the shower, and he builds bots the same way. Each bot is managed by the domain expert who knows that problem best, who treats it like an apprentice. With Team Bots, launched about two weeks earlier, the sales team shares an expert bot everyone contributes to, and the data team handed routine analysis requests in Slack to a 'data bot' with read-only database access.
One Slack thread, 200 messages, one feature
Tan showed a feature she built end to end in a single Slack thread with her colleague Samantha and her bot 'potato.' The bot spawned a Cursor cloud agent to prototype it and posted videos and screenshots back to the thread, where people gave feedback on the spot. The thread ran to about 200 messages, yet Tan said she never opened Cursor for the project and did not read the PR. Ugarte added that because the work happens in public channels, he sees someone from the same thread adopt the workflow the next day.
11,000 PRs a month, 0.5% reverted
At AI Engineer the same week, Tan said she "shipped 3,000 pull requests to production" last month. The team shipped 11,000 in a month, about 227 people contributed at least one PR in October, and the share of reverted PRs went down on a relative basis to about half a percent. Her method is an 'inverse pyramid': invest most in the layers machines enforce, such as architecture and API design, lint rules, types and tests, and add prose such as rules or AGENTS.md last. It took her six months to get there.
Lauren Tan · Engineer, Grok Bot · YouTube · AI Engineer ↗Why they banned 99% of comments
Whenever you correct an agent, Tan said, make it a deterministic check instead of another line in your rules. At Cursor, agents patched an architecture clash with timers and wrote comments justifying the quick fix, and as the pattern spread the app slowed down. So the Grok Bot team banned React's nuance-heavy useEffect in product code and banned 99% of comments, replacing them with lint rules whose error messages tell the agent to fix the root cause. The codebase is the agent's biggest input, she explained, so one bad pattern spreads like weeds.
Lauren Tan · Engineer, Grok Bot · YouTube · AI Engineer ↗Tokens go to verification, not writing code
Code review has shrunk into looking at verification artifacts such as videos and screenshots, she said. Agents run the app themselves, and a 'feature map' of the product from the user's point of view gives them shared vocabulary so they don't have to guess at "the thingy in the left sidebar." The team model-checks cross-device notification flows with TLA+ and uses Lean on its custom virtualization library. Tan said most of her token usage goes to verification and called this "a golden age of software testing," since agents make checks affordable that used to be too expensive.
Lauren Tan · Engineer, Grok Bot · YouTube · AI Engineer ↗When "I have a PR for that" becomes routine
Ugarte said the bottleneck has moved from engineering resources to alignment on the roadmap. In discussions someone often says "I have a PR for that," but merging everything would lead to "a bit of a Frankensteiny product." So while small bugs found in thousands of daily feedback items get fixed by a bot and go out automatically at 3 a.m., changes to the pixels users see every day are dogfooded for longer than feels comfortable. If something gets only mixed internal product-market fit, they don't ship it. In a world of execution abundance, he said, failures will come from intent specification, not execution.
Pushback: bots without EQ, bots with cards, $0 revenue
The pair acknowledged open problems. In a busy Slack thread a bot doesn't know whom to listen to, like an employee with "no EQ," and on permissions their rule is to ask what you would do for a human: scope access narrowly and expand it as trust builds. StartupHub.ai noted the talk gave no figures on how many bots can spend money or how often approvals fail. DataNorth relayed a report that Musk wrote on X on November 1 that Grok Bot cost too much to serve relative to revenue. And Linas Beliūnas wrote that a 72-hour Grok Bot build by three employees merged over 400 PRs but made $0, noting that bots "don't decide what to build, don't make anyone care."
StartupHub.ai · DataNorth · Linas Beliūnas · StartupHub.ai · DataNorth · Linas's Newsletter ↗From accounting degree to AI engineer
Tan grew up in Singapore and studied accounting and finance in Australia. In a financial modeling course she found building the spreadsheets and macros more fun than the finance, started a startup with a classmate, and taught herself to code. Blogging about what she learned caught the eye of people in the Ember JavaScript community, who invited her to speak at her first-ever conference, and she later worked on the React compiler and at Cursor. "Do cool stuff, talk about the cool stuff," she said, closing with the idea that a single tweet can change your life.
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.


