What they did
Simon Willison wrote up a piece by Jacob Sansbury about how his company builds Spawn, a massively multiplayer game where players prompt an in-game agent, Savi, to change the world. By Sansbury's account the game is now built by a swarm of agents running on Anthropic's Fable models that edit the live code and deploy to production every hour. The agents read errors, logs, metrics and bug reports themselves, and about 60% of the work starts from one of those. Savi files bugs on its own when a player or Savi hits one, and an agent decides whether it is a real bug, a won't-fix or a design call. Monitors and log alerts start or wake an agent that owns the incident, and the team breaks things on purpose to test those recovery paths.
How it went
Sansbury says a reported bug usually becomes a fix delivered to that player within an hour, with a reply in the same chat. The team treats an hour with no ship as its most severe incident level and can roll back a broken change in 30 seconds. He credits Fable 5 and 5.1 with making this possible and says Opus 5.5 was not enough. A trial of GPT-6 Astra on security-review fixes went badly: it disabled every insecure feature instead of fixing them. Willison tried the game on his phone and prompted a castle and a dragon into being, but did not check the development claims.
Worth knowing
The token bill is reported as over $3 million a month, and every number here is the company's own claim. Willison doubts this way of working would suit software where failures cost more than in a game. The model is named but the agent harness is not.