Skip to content
pulse

three spacexai employees built a company with ai bots in three days the hard part was management, not prompts

Man drawing on a glowing org chart of robot icons – managing AI agents – as two people work on laptops nearby. pulse

grok bot galaxy was a public experiment, streamed live. @mattyp, @poteto and @roshan_s started with no product, no code and no team. instead of people they hired grok bot – spacexai's ai agents, each with its own cloud computer

- day 1 – the first idea: software for chefs who run pop-up restaurants in san francisco – ticketing, venues, staff
- day 2 – pivot. a real-world event needs city permits, and permits take longer than three days. the team switched to a game
- day 3 – thursday arena is live: a free browser game where teams of bots fight each other automatically. we played it, it works

the workshops spent little time on how to write prompts. they read like a manager's handbook:

- proof before trust. @lingxi's bot cleans up the codebase at 3am every night. its changes go in without human review only if it attaches proof: screenshots for interface changes, speed measurements for performance ones
- stop conditions. @SimonLackowskii's sales bots build a list of 50 potential customers a day, flag 5 to contact now and schedule the rest. once a deal moves forward in salesforce, the automated outreach emails stop
- probation. @davidgan's customer support bot starts by only reading requests, then writes draft replies for a human to check, and answers customers directly last. his own cost estimate: $1-2 per complex request, ~$0.20 for a simple one
- headcount discipline. before adding another bot, ask whether a scheduled task for an existing one would do. if you check on every bot yourself, you become a manager with too many reports
- access breaks before the work does. the bots got stuck on expired logins and files nobody had permission to open, not on the tasks themselves

the takeaway: ai agents can already do the work. the bottleneck is no longer the prompt – it's how you manage them

before you launch a bot, decide three things:
- what it can access
- what counts as proof the job is done
- when it must stop

then widen its permissions step by step, like a new hire on probation

full day 3 recording in the quoted post:

ON AIR · RADIO.THEHYPE.NEWS ↗ ai news radio — 24/7