How SpaceXAI Built GroqBot in One Month

Imagine hiring someone and telling them, on day one, that they will not get a laptop. They will sit next to you and share yours. They will use your logins, and you will use theirs. No company would design a job this way. Yet according to Roman Ugarte, head of growth at Cursor and one of the builders behind the AI agent GroqBot, this is roughly how most people currently work with AI.

"We're onboarding these super intelligent new colleagues, these AI bots, and you're asking them to share the same computer that you have. It's crazy." — Roman Ugarte [32:33]

That observation, half joke and half thesis, explains a lot about why GroqBot, built inside the company now called SpaceXAI, became one of the fastest-adopted AI products of the year. Ugarte was employee number 15 at Cursor, the AI coding tool (01:38), and for the last two years he has run growth there. Around a month before this conversation, a small team inside the company went, in his words, into a cave. Their assignment: build an AI product for everyone who is not a software engineer. From the first line of code to a working internal prototype took only about a month (05:07). Three weeks after that internal release, the product went public. By the time this interview happened, it had been live for three weeks and had already drawn a standing-room-only meetup.

Speed like that is usually explained away with a shrug: good team, good timing. Ugarte's account is more specific, and more useful, than that.

Small rooms, fast decisions

The GroqBot team was deliberately tiny and deliberately cut off. It sat in a separate part of the office, on private Slack channels, isolated from the rest of the company. Ugarte argues this was not a nicety but a mechanism: a handful of people had to make dozens of small, non-obvious product decisions every day, and a bigger group debating a twelve-month roadmap would never have converged. Once the internal prototype existed, the team did something that sounds almost old-fashioned for an AI product: they onboarded people by hand. Two to three hundred early users, one at a time, on calls that ran about twenty minutes each (11:41, 12:05). Some of those calls were rough. A computer would not spin up, or a new user would sit there confused. Ugarte's rule was that the team had to be on those calls, because a painful onboarding today became a fixed bug tomorrow.

The manual-onboarding list was not just power users and AI insiders. It included, deliberately, people the company would never normally talk to, among them a coffee shop owner recruited through a friend of a friend, who became both a devoted user and, in Ugarte's telling, a rich source of bug reports about things like a flaky Shopify integration. The point was to find the blind spots of a team that otherwise lives, as Ugarte puts it, in the Silicon Valley AI bubble.

The onboarding calls also surfaced a pattern the team hadn't designed for. Inside the company, some employees started giving one of their bots a promotion, naming it a kind of chief of staff that would receive instructions and then fan tasks out to a roster of five to ten other, more specialized bots (12:46-13:42). The team noticed the pattern, decided it was worth encouraging in the product, but was careful not to force it on new users before knowing whether it held up outside the company's own culture. It did.

These sessions also taught the team what people did not want to see. Competing AI tools tend to show their work: a scroll of tool calls, a visible chain of thought as the model reasons its way through a task. GroqBot's team went the other direction. A user sends a message, the bot starts working, and the interface shows a simple typing indicator and progress updates as the bot sees fit, nothing more (15:16-15:42).

"We really wanted to take a strong stance that users did not need to see all of those mechanics." — Roman Ugarte [15:42]

Some users asked for a visible to-do list. Nobody asked for the raw stream of reasoning text. That distinction, small as it sounds, shaped a large share of what got cut in the three weeks between internal beta and public launch, a period Ugarte describes as mostly unshipping: removing debug-style visibility tools, developer-y dashboards, and general jank that had crept in while the core team was building for itself (19:42).

A computer of one's own

The decision Ugarte returns to most is architectural. Most AI agent tools available before GroqBot ran on a version of the user's own device, sometimes literally: a home computer left on, tethered through a VPN, so the agent could keep working. GroqBot's team decided early that every bot should run on its own persistent computer in the cloud, separate from the user's device entirely (31:08). That choice does two things. It means a bot keeps the same state everywhere the user reaches it, from a phone or a laptop, without anything needing to stay powered on at home. And it means the bot can use a computer the way a person does, clicking through a browser and typing into forms, for the large number of everyday tools that were never built with a clean API. Ugarte points to the company's own sales team as an example: tools without well-supported integrations had been unreliable for other AI assistants, and only stopped stalling once GroqBot could just operate the interface itself, mouse and all (21:56-23:02).

Ugarte credits an earlier product called OpenClaw, an enthusiast setup that ran an AI agent off a home computer, with proving the underlying idea: give a model access to the tools a person actually uses, plus its own computer, and even today's models get remarkably far (36:32-37:33). What GroqBot added was packaging: no VPN, no Mac Mini in a closet, an experience simple enough that recruiting teams at the company, according to Adam Ward, the head of talent, used it to build always-on sourcing bots that scan academic conference websites every morning for new co-author names and draft outreach messages (25:50-26:10).

The same logic runs through what Ugarte calls the company's internal culture, described with a phrase employees use casually: colleague-pilled. When a product debate stalls, the tie-breaker is not a metrics dashboard but a simple question: what would you want from a human teammate in this exact situation? He offers the number that anchors why this matters to him personally. A teammate you trust only 90 percent is still a teammate you have to watch, which means the work is not actually 90 percent done, it just feels that way while it sits on your mind (62:59).

"An AI that does 100% of the job feels categorically different from one that gets you 90% there. I've significantly updated what I think AI is capable of." — Roman Ugarte [60:43]

That gap, in Ugarte's account, is the real product. Cursor scaled from 15 employees to more than 1,000 (64:33) without, he insists, losing what he calls scramble energy, the willingness to treat every six months as a moment to reinvent the product rather than defend it. He is candid that this is existential rather than aspirational: a company that cannot rebuild its priorities every six months, he says, will lose (68:49). Asked about moats, a word every AI founder currently loses sleep over, he declines to offer a strategy diagram. Cursor's advantages, he argues, were not planned; they were the residue of an obsession with building something genuinely useful today, then deleting it three months later once the underlying models made it obsolete on their own (73:14-74:20). GroqBot, on this account, is not a finished thesis about the future of work. It is what happens when a small team keeps asking, every few months, what a human colleague would actually want, and then goes and builds it before anyone gets a chance to overthink it.