Lenny's PodcastHow we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
At a glance
WHAT IT’S REALLY ABOUT
Building Grok Bot fast: cloud colleagues, hidden complexity, relentless simplification
- Roman Ugarte describes how a small, isolated team built Grok Bot from scratch in about a month, then iterated for three weeks before public launch by focusing on reliability and simplification.
- Grok Bot’s key differentiators are cloud-first execution and giving each bot its own computer, which expands what it can do beyond API-enabled tasks and removes device-dependence and workflow “jank.”
- The team manually onboarded a few hundred early users to surface confusion, validate emergent usage patterns (like a ‘chief of staff’ bot), and gather feedback from non-traditional profiles outside the typical AI power-user bubble.
- A central product philosophy is to hide internal mechanics, reduce UI clutter, and frame progress as “Grok Bot can now…” capability gains rather than shipping more buttons and settings.
- The long-term vision is a true team of AI colleagues for work and life: long-lived agents with memory, proactive monitoring/digests, and bot-to-bot delegation that feels like collaborating with human teammates.
IDEAS WORTH REMEMBERING
5 ideasStarting from scratch enabled a focused, consistent knowledge-work experience.
Rather than extending an existing developer-first tool (Cursor), the team created a clean-slate product with a single coherent vision for knowledge work, avoiding “shipping the org chart” via tabs and clutter. This let them optimize every pixel and interaction for non-technical users and the teammate paradigm.
Cloud-first + “a computer per bot” is the core architectural unlock.
Two pivotal design choices: (1) tasks run in the cloud so workflows persist across devices and don’t depend on the user’s laptop being awake; (2) each bot has its own computer, letting it operate like a real colleague on tools that lack APIs/MCPs. This unlocked reliability and expanded the set of automatable workflows (e.g., sales tools with finicky UIs).
Manual onboarding was a deliberate product-research engine, not a growth stunt.
The team held 200–300 manual onboarding calls over ~2 weeks to observe confusion, failure modes, and real-world use cases outside the Silicon Valley AI bubble. That tight feedback loop drove immediate fixes (“that can never happen again”) and helped validate emergent patterns like the ‘chief of staff’ bot without leading users.
Abstraction beats transparency: users don’t want to babysit the agent.
They intentionally hide most internal mechanics (tool calls, step-by-step clickstreams, chain-of-thought-like streams) to avoid overwhelming users, mirroring how you don’t watch a teammate’s every keystroke. Users wanted higher-level visibility (e.g., a to-do list), but not raw internals—supporting the “colleague” UX thesis.
Unshipping and ruthless simplification is a competitive advantage in AI UX.
A major push from internal beta to launch was removing experimental/diagnostic UI and reducing “Grok Bot now has…” feature clutter in favor of “Grok Bot can now…” capability improvements. Example: automations should be created in natural language ("Remind me at 8am") rather than via trigger/action builders, making “pixels” disappear wherever possible.
WORDS WORTH SAVING
5 quotesOnce 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 Ugarte
We un-shipped a lot.
— Roman Ugarte
And instead to reframe it as, "Grok Bot can now..." Which is, is I think a much more human way of kind of describing these capabilities.
— Roman Ugarte
It's crazy. Like, if you were onboarding someone to your team and you said, "It's your first day. I'm gonna onboard you. Uh, you don't have your own laptop. You're gonna sit next to me. We're gonna share this laptop forever and constantly trip over each other. You're gonna have access to my credentials. I'm gonna have access to your credentials."
— Roman Ugarte
An AI that does 100% of the job feels categorically different from one that gets you 90% there.
— Lenny Rachitsky
High quality AI-generated summary created from speaker-labeled transcript.
