How we built Grok Bot in a month

Roman Ugarte, a leader at SpaceXAI (formerly on Cursor's go-to-market side), explains how the Grok Bot knowledge-work agent product was built in roughly a month by an isolated small team, then iterated and launched. The conversation covers early manual onboarding, cloud-hosted bots with their own computers, the decision to avoid exposing internal mechanics, and broader views on AI teammates, moats, and company culture.

01Key learnings

• Build a new product from scratch with a small, isolated team rather than retrofitting an existing surface, so each micro-decision can be made quickly.
• Onboard early users by hand and sit in on calls; fixing failures the same day compresses learning dramatically.
• Let users reveal patterns on their own before productizing them, then gently encourage the winning patterns without making them one-way doors.
• Hide internal agent mechanics like tool calls and chain-of-thought; users want progress updates, not a stream of model internals.
• Judge every feature by the launch post test: if the sentence 'Grok Bot can now...' isn't compelling, don't build it, and unship jank aggressively.
• Host agents in the cloud with their own computers and persistent memory so they behave like colleagues rather than sessions tied to a local device.
• Design for the 'colleague' frame: ask what you would want from a human teammate in the same situation, and build toward that.
• Find moats by building something people are obsessed with and pulling tomorrow's capabilities into today's product, rather than planning moats upfront.

02Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.