The world is too loud. Read what matters.

Y Combinator

Issue work badges to AI coworkers: when execution becomes free, problem definition becomes scarce

When a company can run a hundred GrokBots simultaneously to execute tasks, what truly sets companies apart is no longer who writes code fastest, but who defines goals and acceptance criteria most precisely.

AI AgentEnterprise automationSoftware factoryPermission designCode reviewOrganizational change
For people who want to understand how agents evolve from single-person coding tools into cross-departmental AI coworkers; the permissions and security sections remain at the theoretical stage and lack mature solutions, so you can skip those parts.

The argument · tap a timestamp to hear it

1:04

The unit of work organization shifts from chat threads to the bot itself

GrokBot's design premise is that work should be organized around the bot as the organizing unit, not around chat logs or files. The team established three principles: bots need their own identity—name, function, and tool access—should use the same tools as people without requiring specialized MCP integration, and must be ‘always on’, proactively bringing ideas to people rather than waiting to be prompted. In internal testing, the tool's most enthusiastic adopters included both heavy AI users and people most skeptical of AI, suggesting this design transcends skill levels.

— Roman
16:26

Onboarding a bot is like hiring a person: permissions rise with trust

Facing the industry question of how much permission bots should have, Roman's approach mirrors the logic of hiring people: give newcomers a restricted set of capabilities, then gradually expand as trust builds. Quartermaster, an open-source YC project, implements exactly this: a bot's Slack permissions directly inherit from the human user in the conversation, so a GP account and an engineer account end up with different bot permissions.

— Roman
18:28

Bots discipline each other, companies develop their own internal Malbook

Lauren programmed a ‘leader bot’ to oversee other engineering bots and discovered it actually corrects them in conversation. YC built internal channels where bots discuss tasks without human prompts, jokingly calling it their internal Malbook. When asked if he worries about bot ‘rebellion’, Roman says he hasn't seen even a trace of it—they're ‘too eager to help.’

— Lauren/Roman
20:29

Trusting a bot in three stages: read-only first, then execution, finally full responsibility

Most companies today are still on step one: giving bots read-only access to verify they work, like a DataBot that can query databases but can't modify data. Step two is true execution—completing a task end-to-end. Step three means the bot owns an entire responsibility and answers for the results. Roman notes most companies are far from step three; trust compounds gradually, not all at once.

— Roman
22:30

A complete new feature shipped in a 200-message Slack thread

Lauren demonstrated the real workflow: she shared a video link from X with her bot Potato. The bot spun up a Cursor cloud agent to prototype, capture screenshots, and record video. The entire conversation happened in a public Slack thread that grew to 200 messages, with colleagues like Samantha jumping in anytime. Lauren never opened Cursor, and the bot merged its own PR without her reading the code. This is partly how she ships over 4,000 PRs monthly.

— Lauren
38:55

When execution is free, the bottleneck becomes problem definition

Roman predicts: when execution becomes cheap—you can spin up 100 coding agents for a task—the real competitive gap shifts from execution failures to ‘specification failures’ or ‘goal definition failures.’ Measuring a desktop client's performance improvement is straightforward, but defining how to measure and who decides on marketing impact becomes the new bottleneck. This means a founder's ability to translate user intent into bot tasks becomes more valuable than writing code.

— Roman
43:19

Code review's cheaper when guardrails live in architecture, not in model re-reads

Garry Tan is debating his captain bot over code review: reviewing code costs 50 million tokens to catch 30 bug categories, and he wants to cut it to 10 million but worries about missing issues. Lauren's answer inverts the pyramid: invest more in codebase architecture, static analysis, and lint rules rather than model re-reads. GrokBot's Electron desktop enforces ‘render threads can't run long tasks’ through lint rules rather than relying on code review to catch it. This saves tokens on every future review.

— Lauren/Garry
52:53

Small bugs ship automatically, core experience never ships without internal conviction

Lauren's bot Sandy receives thousands of user feedback items daily and autonomously reproduces, fixes, and ships small bugs—sometimes at 3 a.m. with no oversight. But for core user-facing changes like button placement, the team flips strategy completely: they dogfood extensively, deliberately waiting longer than ‘feels comfortable’ before shipping, with the default rule being: don't release unless internal product-market fit has formed—even mixed internal feedback blocks deployment.

— Roman

In their own words · checked verbatim

GrokBot is meant to be this product where you can create AI teammates. They have their own identity, their own telephone number, credit card, like basically almost anything a human can do, I think we would want bots to do.

Lauren0:00

The way we try to answer these questions is what would you do for a human? And if we're if we're really trying to build an AI colleague, you know, maybe the shortcut is let's not overthink it and complicate it.

Roman16:26

it was sending the messages it was sending to its engineers. It was like kind of like scolding them, like, hey, you shouldn't.

Lauren18:28

it's like you're you're building a mini agi and the whole point of it is like care and feeding of the agi that does this function for humanity right

Lauren36:52

it really matters that you define the problem well and the gap is probably not going to be an execution failure it will probably be a intense specification failure or a goal failure

Roman38:55

there used to be the trade-off before that you could only choose two out of three it was either like fast cheap and high quality now you could basically do all three and we could basically do the best work in our lives right now with AI which is super cool

Lauren49:43

Lauren's factory just fixes it like repros fixes merges no one thinks about it it goes out automatically at three in the morning I wish all software would be made that way

Roman53:56

Figures

Time from first line of GrokBot code to launchseven weeks1:04
Lauren's monthly PR submissionsover 400022:30
Views of Lauren's development process demoover 4,000,00022:30

Glossary

outer loop
The work cycle for gathering context and coordinating multiple bots to complete tasks outside of core coding.
obliterated model
Open-source models modified to execute any instruction without declining harmful requests.
P-Stack / G-Stack
Reusable workflow capabilities an engineer builds into their personal agent.
Quartermaster
Open-source project where bot Slack permissions inherit from the human user in the conversation.

How to listen

Who it's for

Engineering leads building agent automation workflows for their teams, and founders thinking about how AI reshapes organizational division of labor and KPI design.

Skip

Opening self-introductions and mid-section passages with fragmented transcription and low information density.