How the Most AI Pilled Company in the World Works
Spawn, a 10-person startup, runs a persistent fleet of agents that ship thousands of changes daily; its playbook of closed loops, AI judges and human taste offers lessons for other companies.
Watch on YouTube ↗Charlie's summary
- ▶ 17:11
Spawn runs a persistent, named agent fleet, spending 30 times more on tokens than on payroll
Unlike most agent setups that spin up and down per task, Spawn's fleet is persistent. Each agent has a name, identity, memory and ownership of a function such as the game engine, bug hunting or finance. Spawn spent $3.1 million on tokens last month, 30 times payroll, and shipped 77,000 code changes, about 108 an hour. Jacob says they run a small human team alongside a big team of agents, rather than big human teams that use AI.
- ▶ 19:42
Remove the systems built for human work: Spawn killed pull requests and agent chatter
Jacob found that most problems were solved by deleting processes designed for humans. Agents spent their time telling each other what they were changing, and “double the agents and you roughly quadruple the chatter.” Agents understand the whole system, so Spawn dropped pull requests and let them work directly in the codebase. A broken change spawns a triage agent, and of 123,000 changes in a year only about a thousand were reverted. The cost is heavy compute.
- ▶ 21:45
Close the loop: let players' agents report bugs so the software heals itself
Jacob says the human relaying faults back to agents was a waste of time. In Spawn, players' agents detect bugs and file reports automatically. A small model triages them and routes them to the right agent in the fleet. One issue was detected 8 minutes after affecting a player, named a minute later, and fixed 11 minutes after that. Spawn treats a missing hourly ship as an outage, because a system that moves steadily is more stable than one that freezes and then jumps.
- ▶ 23:16
Turn decisions into precedent: AI judges trained on two years of Jacob's taste calls
Jacob woke to 700 agent questions and realized he had made most of these decisions before. Spawn turned two years of decision logs into principles and trained AI judges. A judge reuses a prior human decision with an audit trail, sends nonsensical questions back, or escalates novel ones. In the first 8 days they made over 11,000 decisions, and Jacob decided only 27. Daily output rose from around 1,000 changes to over 7,000.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.