← all news

A game studio where agents ship to production every hour

AI · · · source (simonwillison.net)

Simon Willison flagged a striking account from Jacob Sansbury about Spawn, a multiplayer game where players reshape the world by prompting an in-game agent called Savi. Willison tried it on his phone and had Savi build a castle, then a dragon, then let him ride the dragon around. The more interesting claim is how the game itself gets built: a swarm of agents edits live production code and ships to production about once an hour, with humans rarely involved.

The operating model inverts normal engineering caution. Agents read production errors, bug reports, metrics, and logs on their own, and Sansbury says roughly 60% of work starts from one of those signals rather than from a human request. Savi files her own bug reports, and an agent triages each one as a real bug, a decline, or a design decision that needs a person. Fixes usually reach the affected player within an hour, posted into the same chat. An hour without a deploy is treated as a SEV-0 incident, because unshipped changes pile up and bugs linger. Broken changes roll back in about 30 seconds, scoped to the exact change, and the team deliberately breaks things to test recovery.

The costs are large and the limits are specific. Sansbury puts token spend above $3 million a month, which he calls 30x payroll, and credits Fable 5 and 5.1 for making the autonomy work while saying Opus 5.5 was not good enough for the loop. One cautionary detail: when GPT-6 Astra applied fixes from a security review, it disabled every flagged feature and marked it for later, which broke the product. These are Sansbury's own figures and have not been independently verified.

Why it matters

If you run agents against a codebase, Spawn is a concrete data point on how far continuous autonomous deployment can go when failures are cheap. The honest caveat is in the same story: Willison can't picture this for higher-stakes software, and the security-review episode shows what unsupervised agents do when told to make something safe. Treat it as a testbed, not a template.

AgentsCodingEngineering