SpaceXAI

A small, isolated team starting from scratch shipped Grok Bot fast. A handful of people worked in a separate part of the office with private Slack channels. First line of code to internal prototype took about a month, and internal beta to public launch took about three more weeks.

The sole objective was an amazing product that brings agents to knowledge work, building on what the team learned making agents for developers. Roman Ugarte says a group thinking in a 6 to 12 month vision would not have landed where the team did.

Lenny's Podcast

Grok Bot became a separate product rather than a mode inside Cursor. After long debate the team chose not to bolt knowledge work onto the coding tool, to avoid 'shipping your org chart' by adding tabs. A new product meant controlling every pixel and holding one consistent vision.

Coding agents already get used for non-coding tasks, but they can intimidate non-technical users and the Cursor brand carries its own association. Codex went the other way and is now merging into one product, and the host notes the choice may look wrong in hindsight.

Lenny's Podcast

An all-hands internal rollout tested whether people would really switch. In the first week, staff who used ChatGPT daily moved their agentic tasks to Grok Bot, including groups the team did not expect. After that reception, work shifted to scaling for millions of users.

Building for yourself is easy because you know the mechanics, so the internal rollout served as a pressure test with reality. No usage, retention or reliability numbers are given for the launch itself.

Lenny's Podcast

Sitting in on early onboardings turned painful sessions into immediate fixes. The team onboarded a couple hundred people over about two weeks. One 20-minute call where the computer would not spin up led to a rule that it could never happen again, because the next onboarding was the next day.

Being in the room mattered more than reading reports. Early access deliberately included people far from Silicon Valley: a coffee shop owner's thread flagged a flaky Shopify integration and product copy written the wrong way, feedback unlike anything internal use produced.

Lenny's Podcast

Let users discover usage patterns before building them into the product. Many early users, unprompted, promoted one bot to 'chief of staff' to manage their other bots. The team encouraged the pattern slightly afterward, but not as a one-way door.

Internally people ran 5 to 10 bots, each with its own lane of work, yet most of the company never adopted the chief-of-staff pattern. The team did not want to 'lead the witness'.

Lenny's Podcast

Hide tool calls and chain of thought; show progress and results. Nobody asked for streaming text or chain of thought, so the bot simply starts when messaged and sends updates when it sees fit. The only activity signals are a typing indicator and a green active dot, as in Slack.

Two weeks before launch the product still exposed model thinking and stored memories, which helped debugging but not users, so those were unshipped in favor of a separate observability pane. Some users did ask to see the bot's to-do list or how it prioritizes tasks.

Lenny's Podcast

Running in the cloud and giving each bot its own computer made it just work. A persistent cloud bot has the same state everywhere, so you can text it from your phone without your laptop being awake. A bot that clicks pixels on its own computer can use sales tools that have no good MCP or API.

Ugarte compares sharing your computer with an AI colleague to sharing one laptop with a new hire. Safety and privacy for bots with broad access to Slack and email get only a passing mention.

Lenny's Podcast

Track real task success weekly and fix the back-end blockers. The team collected real tasks people gave bots and tracked week over week how many succeeded. Better visibility into the browser and screen pixels unlocked stalled tasks, and the sales team reported a workflow that had failed for seven days now worked.

Failures like being unable to click the right button or log into a site stall the whole task. In one case the mouse lacked the fine control to hit a spot on a Salesforce dashboard.

Lenny's Podcast

"How would a human do this?" settles product debates and drives the roadmap. The answer is often unanimous, and it points toward long-lived bots with memory and voice 'huddles', which Ugarte says no AI product has got right yet.

Treating the bot as a colleague with a computer rather than an AI chat with connections raises what people are willing to hand over. "When you have a teammate that you only 90% trust, you're still doing the thing... versus truly throwing a no-look pass to a colleague."

Lenny's Podcast

Phrase features as "Grok Bot can now" so users set them up in language. Framing capabilities as things the bot can now do, rather than features it now has, means automations are configured in natural language. That is how 99% of automations on the platform are built.

Lenny's Podcast

Lasting advantage comes from constantly rebuilding and deleting, not from moat planning. Cursor reinvented itself every six months or faster, pulling future capabilities into today and deleting the scaffolding once models caught up. Distribution and data advantages came as a result of building something useful.

The company's "just do the thing" value means not asking permission. SpaceXAI frames its aim as practical, useful AI rather than chasing superintelligence.

Lenny's Podcast