SpaceXAI just released a free Grok Bot workshop
54 minutes on running a team of bots like a real product org:
03:37 - it gets really hard to babysit multiple agent threads. everyone built their own coordinator agent, never first-class
13:32 - map it to a real company: specialists, and knowing who to turn to with a data question vs a design question
23:59 - agents multitask, several questions in parallel
30:08 - context gets shared behind the scenes, nobody copies and pastes
40:40 - the em agent learned to task out work and review it. nobody told her to
Every routing decision inside that org comes back as a paragraph from a frontier model
Who takes it, what to check, who signs off, all of it as prose
That is not intelligence architecture
Jev returns typed options plus a probability, $0.042 per million input tokens, and never writes a sentence
LLM creates the work โ Jev decides what happens next โ code enforces it
Save this before you build another coordinator agent