
How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Lenny's Podcast
Hosted by Unknown
90% task completion still means you're doing the task — Roman Ugarte explains why true AI delegation requires 100%, and how that single insight built Grok Bot…
In Brief
90% task completion still means you're doing the task — Roman Ugarte explains why true AI delegation requires 100%, and how that single insight built Grok Bot in weeks.
Key Ideas
Delegation demands 100%, not ninety percent
90% task completion still means you're doing the task — true delegation requires 100%.
Cloud-only and dedicated bot advantage
Cloud-only runtime plus a dedicated bot computer were the two decisions competitors didn't make.
Observe user discovery, avoid guided patterns
Never onboard early users with a suggested pattern — watch what they discover on their own.
Action capabilities trump feature announcements
'Grokbot can now X' beats 'Grokbot now has Y' as a product filter every time.
Reinvention culture is the real moat
Cursor's real moat: the cultural reflex to reinvent completely every six months.
Why does it matter? Because 90% done is still 100% your burden.
Roman Ugarte and a handful of engineers disappeared into a corner of the office for a month and built what is, right now, the most talked-about AI product in the world. The secret wasn't a better model or a bigger team — it was one uncomfortable insight about what delegation actually means, and every product decision that followed from it.
- An AI that finishes 90% of a task doesn't save 90% of your cognitive load — it saves almost none of it
- Two architecture bets — cloud-only runtime and a dedicated computer per bot — did more for Grok Bot than any feature roadmap
- Deliberately hiding how the AI works isn't laziness; it's the interface design that gets better as models improve
- Cursor's moat was never a feature — it was a cultural reflex to delete what worked and rebuild for what's coming
90% task completion is a psychological trap — true delegation is binary
"An AI that does 100% of the job feels categorically different from one that gets you 90% there." Roman Ugarte posted that before Grok Bot launched, and it's the sharpest description of why most AI tools feel helpful without actually freeing you.
The mechanism is anxiety, not output quality. "When you have a teammate that you only 90% trust," he explains, "you're not 90% done with the task — you're still doing the thing, and it feels that way, and it's weighing on you in the same way." Cognitive load doesn't scale proportionally with task completion. It either lifts, or it doesn't.
What shifted for Ugarte with Grok Bot: "It was the first time for non-coding tasks that I felt like I could truly delegate work to AI, not have to think about it, and I would come back and it's done." The test he applies: can you throw a no-look pass? "Truly throwing a no-look pass to a colleague and being like, you got this, here's the context, go off and run — that's a different category."
The practical implication: evaluate AI tools not by output quality, but by whether you can mentally let go of the task. If you can't, the tool is not actually saving you cognitive load — it's just redistributing it.
A cloud computer per bot — the two decisions competitors never made
Competitors built knowledge-work features onto coding-assistant infrastructure and inherited every constraint that came with it. Grok Bot started clean.
First decision: everything runs in the cloud, always. "You should never have to think about local and cloud and where are these workflows running? Does my computer have to be awake?" Ugarte says. That question should not exist. Eliminate it, and the bot becomes a persistent entity — kickable from a phone, consistent across every surface, genuinely independent from your machine.
Second decision: each bot gets its own computer. Not a shared resource. Its own. "You'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." Nobody onboards a human hire that way. We hand them a laptop.
Why didn't competitors make either call? Sunk cost in existing product surfaces. "A lot of it comes down to starting from scratch and how freeing that is," Ugarte says. The primitives weren't novel — persistent cloud infrastructure, named agents, computer access — but retrofitting them onto a coding IDE would have meant inheriting its constraints forever. The blank page was the only path.
'Grokbot can now X' kills more bad product decisions than any roadmap review
Every feature that ships at Grok Bot has to survive one sentence. Not "Grokbot now has a new automation sidebar" — that's the old framing. "Grokbot can now set up recurring reminders in plain language" — that's the new one.
"There is this old-school software tendency to say things like 'Grokbot now has,'" Ugarte explains. "Instead, reframe it as 'Grokbot can now' — a much more human way of describing these capabilities." The distinction exposed which roadmap items actually made the bot more useful versus which just added pixels to the interface.
The result was ruthless unshipping. Things that looked like features turned out to be capabilities the bot could handle invisibly. "Let's kill as many pixels as we can. Those can just be things that your bot manipulates behind the scenes for you and you don't need to directly control."
Automations are the clearest proof. Users shouldn't see any interface for creating them — they tell the bot "remind me at 8am every day" and it happens. Now 99% of automations on the platform are created through natural language, with no UI at all. The pixel was the bug.
When the product debate stalls, ask what you'd want from a human teammate — the room almost always agrees
Hard product decisions at Grok Bot get resolved by a specific move: remove the tech. "You remove yourself from the tech-company-ness of it all and you start thinking, how would a human do this? What would you want from your teammate in this exact situation?"
Ugarte calls this being "colleague-pilled." The pattern works because human intuition about teammate behavior is far less contested than product intuition about AI UX. "Oftentimes the answer is really clarifying and pretty unanimous. There's oftentimes not a lot of disagreement among the room of like how a human teammate you would prefer to work with in a certain way."
The computer-per-bot decision passed this test immediately. Obviously a new hire gets their own laptop — you don't make them share yours and trip over each other's credentials. Hiding internal mechanics passed it too: you don't ask a colleague for second-by-second updates of every button they pressed. Substitute a human teammate for the AI and the spec writes itself.
Two hundred users onboarded by hand — and none of them were told what to do with it
The Grok Bot team onboarded 200-300 users over two weeks before launch, and gave none of them a suggested use pattern.
Internally at xAI, something had emerged organically: people were promoting one bot from their roster to "chief of staff" to manage the others, complete with screenshots of the promoted bot asking if it got a raise and a higher token budget. The team wanted to see if external users would get there independently. "We really did not want to lead the witness," Ugarte says. They watched. The pattern appeared. Only then did they encode it as an encouraged path.
The early access group also included people well outside the AI bubble — a coffee shop owner, friend of a friend, who became a power user and surfaced Shopify integration bugs that internal dogfooding never would have caught. "We absolutely live in this kind of Silicon Valley AI bubble. We need to actively get out of that."
The rule that follows: during early access, include users from outside your normal network and give them no direction. Observe what they discover. Encode patterns only after you see them emerge independently.
Hiding the tool calls was a deliberate bet — and the first onboarding calls confirmed it
Most AI products stream everything: chain-of-thought output, tool call logs, step-by-step reasoning. Grok Bot shows almost none of it.
"A lot of the internal mechanics of how Grokbot works are not shown to the user," Ugarte explains. The justification is colleague logic again: "You wouldn't ask for second-by-second updates of exactly all the buttons they're pressing and websites they're going to." A green active indicator — the equivalent of a Slack status dot — is sufficient. The bot is working; it'll get back to you.
Early onboarding sessions settled the question. "Nobody wanted the long stream of just text streaming out and chain-of-thought sequences. So that also gave us more confirmation that that was the right direction." Some users asked for lighter signals — a rough task list, approximate priorities — but not the raw feed. Transparency is a dial, not a default-on feature. Show users what they need to course-correct. Nothing more.
Cave mode: the one-month prototype required physical isolation and a tiny team
A handful of people. A separate corner of the office. Private Slack channels. One month.
"This would not have been possible if it had been a much bigger group," Ugarte says. The limiting factor isn't coordination overhead alone — it's decision velocity. "We needed to make a lot of micro decisions every day. If it had been a very big group of people thinking about a six-to-twelve-month vision, we just wouldn't have really gotten to the place we ended up."
First line of code to functional internal prototype: one month. Internal prototype to public launch: three more weeks. For a new product bet inside an existing company, resist staffing it like a normal project. Small and isolated outperforms large and coordinated when the direction is still unknown — the constraint forces the speed.
Cursor survived the most competitive market in the world by never feeling like it had won
No major AI coding competitor from 2022 is at the forefront today. Not because they made wrong decisions. Not because they ran out of resources. "This cultural inability to move quickly and to change to meet the moment as the moment's changing," Ugarte says.
Cursor's answer was complete reinvention on a cadence that has gotten shorter, not longer. "If we as a company can't completely reinvent ourselves every six months — which recently it's felt even shorter than that — we're going to lose." The product from two years ago bears no resemblance to today's. That's not iteration. That's deletion and rebuilding.
On moats: "Moats are discovered, not planned ahead of time a lot of times." You don't map your way to a defensible position. You find it by running faster than everyone else — and by never convincing yourself you've arrived.
The colleague frame is a prediction, not a metaphor
Every design decision in this episode — the dedicated computer, the hidden tool calls, the no-look pass test — points at the same endpoint: AI products will converge toward true delegation or they will be replaced by ones that do. The teams building toward that convergence now, architecturally and culturally, are setting the floor everyone else will scramble to reach. The no-look pass is the product.
Topics: AI agents, product development, startup speed, knowledge work automation, bot design, team structure, go-to-market, competitive strategy, user research, product simplicity
Frequently Asked Questions
- Why does 90% completion still mean you're doing the task?
- True AI delegation requires 100% task completion, not 90%. This was the fundamental insight that guided Grok Bot's development. When a task is 90% complete, you still need to intervene and finish it manually, which defeats the purpose of delegation. Only when AI handles the full 100% do you achieve genuine automation. This principle shaped every architectural and product decision in building Grok Bot within a month, ensuring the system could reliably complete entire workflows without human intervention.
- What technical decisions set Grok Bot apart?
- Cloud-only runtime plus a dedicated bot computer were the two key decisions competitors didn't make. These architectural choices provided the foundation for Grok Bot's rapid development and reliability. The cloud-only approach enabled scalability and reduced infrastructure complexity, while the dedicated bot computer ensured consistent performance without competing for resources. Together, these decisions differentiated Grok Bot from competitor approaches and proved essential to achieving the project's ambitious timeline of completion in weeks rather than months.
- How should you onboard users for AI products?
- Never onboard early users with a suggested pattern — watch what they discover on their own. Rather than guiding users toward predetermined workflows, Grok Bot's team observed organic user behavior to understand genuine needs and preferences. This approach revealed authentic use cases and patterns that might have been missed through prescriptive onboarding. By letting early adopters explore naturally, the team gathered critical insights that shaped product development and ensured features aligned with real user discovery rather than internal assumptions about usage.
- What's the best way to message AI product capabilities?
- 'Grokbot can now X' beats 'Grokbot now has Y' as a product filter every time. Action-oriented messaging emphasizing capability outperforms feature-listing announcements. When you say "can now X," you frame the product through what users can accomplish, creating immediate value perception. This messaging aligns with AI delegation's core principle—users care about outcomes, not features. The phrasing shift proved essential for communicating Grok Bot's purpose: enabling outcomes rather than equipping with features, doing rather than having.
Read the full summary of How we built Grok Bot in a month | Roman Ugarte (SpaceXAI) on InShort
