SpaceXAI built Grok Bot from its first line of code to a working internal prototype in about a month, with a small team deliberately isolated from the rest of the company.
Roman Ugarte, an AI engineer at SpaceXAI who helped incubate the product, revealed this in a conversation with Lenny Rachitsky, creator of Lenny’s Podcast.
Ugarte said the team was created as a “blank page” exercise to build an AI product for general knowledge work.
A handful of employees worked in a separate part of the office and used private Slack channels, letting them make rapid product decisions without the coordination required for a larger project.
The team then spent roughly two weeks manually onboarding 200 to 300 early users.
Ugarte said core team members sat through individual sessions, watching users encounter problems such as computers failing to start or people becoming confused during onboarding.
The process also exposed unexpected usage patterns. Employees began creating multiple specialised bots, with some using a higher-level “chief of staff” bot to delegate work to other bots.
The biggest improvements before launch were largely invisible to users.
Rather than adding more features, the team focused on making the underlying computer-use infrastructure reliable.
Sales staff, for example, were using systems that lacked reliable APIs or MCP integrations. Giving bots their own computers allowed them to interact with those systems through browsers, clicking buttons and entering information like a human user.
Ugarte said the team repeatedly fixed small infrastructure problems, such as giving agents better browser visibility or more precise mouse control. These changes could unlock entire workflows without requiring a new model capability.
The product was also designed around two early architectural decisions: bots would run in the cloud, and each would have its own computer. Ugarte argued that AI agents should not share a user's machine, just as human colleagues normally have their own computers.
That architecture has enabled bots to work with other bots. Ugarte said he uses a QA bot that can run Grok Bots inside its own computer to test software releases and record results. The team's longer-term vision is for users to have teams of persistent AI bots that can delegate work among themselves.
“The ultimate vision of Grok Bot is incredibly simple,” Ugarte said. “You should have a team of AI bots that help you with your job and help you with your life.”
SpaceXAI launched Grok Bot last month as a beta AI agent designed to function as an always-on digital teammate.
The bots have their own cloud-based computers, allowing them to log into the tools, websites and apps people already use and complete multi-step tasks without requiring dedicated APIs or pre-built workflows. Users can assign work from desktop or mobile, and the bot returns when the task is complete or needs approval.
At launch, Grok Bot was available in beta to selected SuperGrok and Cursor subscribers on desktop and iOS, while enterprise access was placed behind a waitlist.
- Coforge Revenue Rises 54% as AI Engineering and Deal Bookings Gain Momentum
- Tech Mahindra Partners With SpaceXSI to Bring AI Into Software Engineering
- Infosys AI Revenue Rises 29% in Q2, FY27 Growth Guidance Cut
- LTM Launches BlueVerse Industrial AI Lab in Chennai
- Chariot Launches Speech Recognition Model for 10 Indian Languages
- Meta Expects Up To $145 Billion In Capex As AI Investments Expand

