How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
with Roman Ugarte
8 Sep 20265 min read1h 25m
TL;DR
Roman Ugarte's small SpaceXAI team built Grok Bot — now arguably the hottest AI product in the world — in just one month by making two non-obvious early decisions: starting a completely separate product rather than extending Cursor, and hiding internal agent mechanics from users to make bots feel like colleagues rather than tools. The internal rollout at SpaceX organically produced a 'chief of staff bot' pattern, and 200–300 hand-onboarded early users validated that ruthless simplification — unshipping features, not adding them — was the real unlock.
Key Moments
Roman Ugarte
“We unshipped a lot. I wish I could have shown you what things looked like maybe two weeks out from launch where I think we'd realized that the core team, we had a lot of experimental features that we wanted to get internal feedback on.”
Roman describing what changed most during the three-week internal beta — cutting features, not adding them.
“I want you every morning to go to the conference website, download the PDFs. There are any new ones. You should find every new name that we've not yet tracked. You should add that name to a spreadsheet. You should do research.”
Roman describing an always-on recruiting bot that the SpaceX talent team built to source candidates from academic conference papers.
“Grockbot now has — and when you think of completing that sentence it would be like a new button to press — instead to reframe it as Grockbot can now, which is a much more human way of kind of describing these capabilities.”
Roman explaining the internal framing shift from shipping UI features to shipping agent capabilities.
“You should tell your bot, remind me that at 8 a.m. every day, please. And then it should just do it. And you should never ever have to see that interface of creating an automation.”
Roman contrasting Grok Bot's natural-language automation setup against competitors' clunky sidebar UI flows.
“Once you start breaking out of this is AI chat with a set of connections instead to this is a colleague with a computer. It just raises the ceiling of what you would think to give to AI.”
Roman explaining the core mental model shift that makes Grok Bot feel categorically different from other AI chat products.
▶ 0:00
About Roman Ugarte
›
Roman Ugarte was employee number 15 at Cursor before joining SpaceXAI, where he led the incubation of Grok Bot from first prototype to public launch. He has spent the last two years in growth and product roles, building intuition for agentic AI products for both developers and knowledge workers. Roman leads product for Grok Bot and was part of the core team throughout its rapid one-month build and subsequent launch.
Takeaways
1
Always-on bots unlock recruiting work humans can't scale SpaceX's recruiting team built Grok Bot workflows that monitor academic conference websites daily, extract author names from PDFs, cross-reference them against internal networks, and auto-draft Slack introduction requests. This kind of always-on sourcing was previously too manual; now it runs without human intervention.
2
Hide agent mechanics; treat bots like colleagues Grok Bot deliberately does not expose tool calls, chain-of-thought, or click-level logs to users. The design bet is that as models improve, users want outcomes — not transparency theater. Early onboarding sessions confirmed nobody wanted a long stream of agent steps; they wanted a typing indicator and a result.
3
Frame roadmap as 'can now,' not 'now has' Internally, the team reframes every potential feature from 'Grok Bot now has [button]' to 'Grok Bot can now [do thing].' This forces product thinking away from UI additions and toward genuine agent capability expansion — a small linguistic shift with large prioritization consequences.
4
Manual onboarding of ~200 users reveals blind spots fast The core team personally onboarded 200–300 early users over two weeks, including unconventional profiles like a coffee shop owner. This surfaced non-developer use cases (Shopify copy, small business workflows) and validated or invalidated internal product theses — like the 'chief of staff bot' pattern — without leading the witness.
5
Start fresh instead of extending existing products The Grok Bot team explicitly chose not to build inside Cursor, recognizing that bolting a knowledge-work agent onto a coding IDE would produce a cluttered product that 'ships your org chart.' Creating a separate surface with a single consistent vision was a major driver of early product clarity and user love.
6
Unship aggressively before launch The core team spent much of their three-week internal beta cutting features, not adding them — removing developer-facing debug UIs and experimental toggles that had crept into the product. Roman's heuristic: if a capability doesn't need pixels, kill the pixels and let the bot handle it silently.
7
Small isolated teams ship breakthrough products faster The Grok Bot prototype went from first line of code to internal launch in one month with a handful of people in a separate part of the office with private Slack channels. Roman attributes this speed directly to the volume of daily micro-decisions that would have stalled in a larger group.