A team building the multiplayer game Spawn says a swarm of AI agents now changes its live code and ships updates hourly, with broken changes rolled back in 30 seconds. The account offers a concrete, unusually ambitious look at agent-led software development in a game environment, including the operating practices meant to keep it moving.
Watch Desk analysis
What happened
In an article published on 29 October, Simon Willison recounts Spawn founder Jacob Sansbury’s description of the system. Players can use the in-game agent Savi to change the world; the development team also uses Fable-powered agents to inspect errors, bug reports and metrics, decide what needs fixing, and make changes. Sansbury says the agents cost more than $3 million in tokens per month.
According to Sansbury, agents can roll back a broken change in 30 seconds, and alerts can wake an agent responsible for an incident. The team deliberately breaks things to test its recovery paths. Willison says Fable 5 and 5.1 enabled this level of autonomy for the team; he also describes a trial of GPT-6 Astra security fixes that disabled features rather than fixing them.
Why it matters
This is more than an AI feature inside a game: the reported development process gives agents access to production code and a rapid route to shipping changes. That makes monitoring, incident ownership and rollback part of the story, not backstage housekeeping. It is also a strikingly expensive experiment, at the monthly token spend Sansbury reports.
The account is one company’s description, relayed by Willison, not an independent audit of the system’s reliability or costs. Still, it gives developers specific practices to debate: short deployment cycles, targeted rollbacks and agents that respond to production signals.
Our read
Spawn is an intriguing test bed for agentic engineering, where mistakes can be caught and reversed quickly. The 30-second rollback is the detail that makes hourly shipping sound less like a dare and more like an engineered process. But the model that works for a game may not travel neatly into software where a bad change carries a much higher price.
What to watch
- Whether the team shares results on how often agent-made changes cause incidents or need rollback.
- How the reported monthly token spend changes as the system scales.
- Whether other teams can reproduce this pace without accepting the same risks or costs.
Discussion spark: Would you let AI agents ship code directly to production if rollbacks were this fast, or should a human approve every change?
Sources and evidence
- Building a multiplayer game world with autonomous Fable-class models (29 October 2026, 17:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.