Grok Bot, a tool for creating AI teammates, went from its first line of code to launch in seven weeks, according to Y Combinator. In a Lightcone episode uploaded November 14, 2026, core-team members Lauren Tan and Roman Ugarte describe their day-to-day workflows: bots working in Slack, coding agents checking their own output, and humans deciding which ideas deserve to become features. Source Source
Ugarte describes building an internal prototype in about seven weeks, releasing it internally and then bringing it to market. Neither account gives exact dates for those milestones. Source Source
The builders describe Grok Bot as a way to create AI teammates with names, identities and access to work tools, designed to work without waiting for each new prompt. Ugarte says their performance-marketing team gave a bot its own login, access to Google Ads and Facebook Ads, and a budget, with instructions to coordinate with a design bot on assets for A/B testing. He doesn’t provide the budget amount or campaign results. The coworker pitch has reached the expense-account stage. Source
One Slack thread, not one magic prompt
Tan says she built an entire feature with her colleague Samantha, another product manager and her bot in a single Slack thread, without opening Cursor herself. That thread contained 200 messages. “One thread” is doing a lot of work there. Source
Cursor technology was still involved: Tan says her bot spawned a Cursor cloud agent, which could run the application and return prototype videos and screenshots. Keeping the discussion in a shared channel, she explains, let colleagues contribute feedback as the work developed, rather than waiting for her to return with a finished implementation. Source
Reaching that point required investment in agent instructions, codebase constraints and automated code checks, Tan says. She now lets agents merge their own pull requests—proposed code changes—and says she didn’t read the pull request for that particular feature, instead assessing the resulting software. Source
Tan recommends having agents run applications and examine results so they can find and fix problems. But she cautions that the loop doesn’t guarantee good code: a working implementation can still be slow or inefficient. For critical infrastructure, databases and deployment, she says engineers may still want to remain in the loop. Source
Who gets access, and what gets shipped?
Ugarte advocates a gradual approach to permissions: start with limited, read-only work, assess the results and expand access as trust develops. He says there is no industry-wide silver-bullet answer to the permission problem, and describes bots owning whole job responsibilities as something still quite far away. A bot getting an ad budget isn’t an argument for handing it every password.
Comments