SpaceXAI lead engineer (Cursor & GrokBot) Lauren Tan: "sometimes i actually don't even look at the PR until after it's landed." in a 44-min interview, she and lead designer Peng Zheng show the 14 bots the GrokBot team actually runs for work and life: 1:02 → a chief of... staff that buys, lists and reprices gear on its own 5:24 → PM, design and eng bots debating one idea in a group chat 16:52 → Dr. Eggbot: a bot that designs your bots and audits them 24:15 → an eng lead bot that never codes, only delegates to agent swarms 37:52 → how to trust bots with more: correct once, make it a skill, then a routine how to build an AI crew that learns by watching you work – pinned article↓show more

bro im so tired of her and this marketing already

fair, but the eng lead bot that's forbidden from coding is a real pattern worth stealing, whoever it comes from

@0xShoopy if the PR process is streamlined, maybe it's already doing most of the work for you. I realized this when I hit too many bottlenecks building my own AI tools.

exactly, fix the pipeline once and every PR after it gets cheaper to review

@0xShoopy nailed it. what’s your go-to for fixing pipelines?

the PR line got me. i used to read every diff before merging, now the tests catch more than my eyes do and i review after it lands.

same shift here. once the checks are good enough, reading every diff before merge is mostly habit

this matches what i see. my read time now goes to design choices, the tests catch the actual breaks.

only works if the team trusts the bots to handle their tasks efficiently

and that trust is earned in steps: watch it once, correct it, save a skill, then put it on a routine

how do the bots handle feedback from PRs if you don't look at them beforehand?

the checks run before anything lands: sub-agents on different models test the change against the goal, and only then it merges. she reviews after, the bots catch most of it before

The interesting shift is trusting agents to handle routine work while keeping a human approval queue. Autonomy only scales when review stays lightweight and the boundaries are clear.

lightweight review is the key. one tap to approve, and everything irreversible still waits for you

the “correct once, make it a skill” loop is the real unlock that’s how agents go from one-off assistants to systems that actually compound

exactly. a one-off assistant resets every chat, a skill turns every correction into something permanent

ai crew sounds kinda like my dream team

the eng lead that's banned from coding would be the first hire on mine

lol facts that sounds like a legend

"Review after merge" is either elite confidence in the setup or a very efficient way to discover production has opinions.

lol it only works because the codebase was redesigned to be agent-friendly first. without that it's the second option

14 bots is actually insane 😭

and each one has a single job, that's why it works

Lauren trusts the AI reviewer to catch errors before merge. This flow saves time and keeps code clean without manual checks.

close, it's mostly evals and tests that gate the merge, not a single AI reviewer. but yeah, the time saved is massive

the wildest part isn’t the 14 bots. it’s the eng lead bot that never codes. it delegates, watches the swarm, and only steps in when the system needs a decision. the human role is moving from doing the work to designing how the work gets done.

Turning a correction into a reusable skill seems like the key step. How do you distinguish a lasting rule from a one-off exception before the bot makes it part of its routine?

a bot that designs and audits other bots is wild, this one's a gem

right? Dr. Eggbot reading every other bot's chats and trimming their routines is the part i'd steal first

wow these bots are next level what's the article link

it's the pinned one on my profile, quoted right under this post

ok thanks for clarifying ill check it out

The luxury of unlimited cursor agents with a crazy verification system. Us mere mortals can't do this without dropping 5k in API credits