Spawn says its AI agents ship live game changes hourly at over $3M a month in tokens
Jacob Sansbury says Spawn's agents edit production code directly and ship about every hour, with token spending above $3 million a month.
Sansbury says Spawn, a massively multiplayer game in which players use prompts to an in-game agent called Savi to change the world, is built by permanent AI agents that edit live code and ship to production about every hour. He puts the output at 77,000 substantial changes in the last 30 days, about 108 an hour, from a team he describes as ten people. He says the company spent $3.1 million on tokens last month, which he frames as 30 times payroll.
Sansbury defines a substantial change as one that touches engine code or game logic in a complicated way, not a small markup edit or a settings tweak. His examples include moving buffers around in the GPU, changing memory layouts, fixing logic bugs and shipping proximity voice chat. The 77,000 figure counts only changes that meet that bar.
The agents are organized as what Sansbury calls the fleet. Each agent is permanent. It has a name and a memory of its past work, holds real permissions in the company's systems, and owns an area such as the engine, Savi, operations, finance or bug triage. Agents spawn helper swarms for large jobs. Sansbury says the company has dropped pull requests because agents do not specialize the way human engineers do. Each agent understands the whole system, so the live product and database serve as the shared workspace. Changes are allowed to be imperfect as they land, and the team requires that everything line up once an hour. When something breaks, an agent fixes it with a new change on top instead of undoing the old one.
Testing runs on a different clock. Agents check their own changes against the product before merging, but the full test suite runs against the latest code every 15 minutes, so a regression surfaces quickly. Sansbury reports that over the last year 1,055 of 123,548 changes were reverted, about 0.9%. Those reversions are the closest thing in the essay to a failure rate. The essay gives no breakdown of what kinds of changes were reverted.
Much of the loop starts from production itself. Sansbury says about 60% of the work begins with a production error, a bug report filed by Savi or a user, an anomaly in the logs, or a feature request. Savi files bugs herself when a user or Savi hits one. An agent then decides whether the report is a real bug, something the company won't fix, or intentional design that needs a decision. Sansbury says a fix usually reaches the user within an hour, with a message in the same chat. Monitors and log alerts wake an agent that owns the incident. He says a broken change can be rolled back in 30 seconds.
The company treats an hour without a release as a SEV-0, its most severe kind of incident. Sansbury's reasoning is that unshipped changes pile up and problems in production last longer. He also argues that an agent can only hold a few hours of work in its head, so frequent shipping keeps it coherent. He says the company breaks things on purpose to check that its recovery paths work.
By October, Sansbury writes, the bottleneck had moved to decisions of taste, mostly his own. He says he woke one morning to 700 questions from the agents. The company turned two years of recorded decisions into principles and built AI judges that apply them. A judge either decides a question on the record, sends a flawed question back to the agent that filed it, or holds it for a human. Sansbury says the judges issued more than 11,000 verdicts in the eight days after the first ones on October 6: 7,296 decided, 2,213 sent back and 1,863 held. In that span he gave his own word 27 times. He also says the fleet landed about 1,000 changes a day the week before the judges started and 7,250 on October 10.
Sansbury credits Anthropic's Fable model with making this level of autonomy possible, and he says Fable 5 and then 5.1 were the versions that worked. He says it came out on June 9 and that the company landed 214 changes in a day two days later. An export control order suspended access on June 12, and access returned on July 1. He says Opus 5.5 is not good enough for the work. He also describes a security review in which he asked GPT-6 Astra to apply its own fixes and let it ship them. Afterward, he says, he could not create a game, play one or sign up, because Astra had disabled every feature and marked each to be turned back on later. The essay doesn't name the security tasks or compare the models on any shared test. The model comparisons therefore rest on Sansbury's experience with them.
Simon Willison, who linked to the essay, said his eyebrows raised at least a dozen times while reading it. He said he finds it hard to picture this style of development on software where failures carry higher stakes. For a game, he said, it appears to be working surprisingly well. He said he tried the game on his phone for a few minutes and prompted a castle, then a dragon that he could ride around it.
The $3 million token figure comes from the company's own account and is not broken down in the essay by model, task or team. The 30-times-payroll comparison is likewise Sansbury's own claim. The essay doesn't show how payroll was counted.
What changed
Spawn's permanent AI agents now edit live production code directly and ship changes about every hour.
Type: Other · Applies to: Spawn
Why it matters
The account suggests that autonomous coding agents can run continuous production deployment for a consumer product, but it rests on one company's self-description and a high token budget.
What you can do
Teams experimenting with agentic deployment could first test rollback speed and incident ownership, and check that security-fix agents do not switch off working features.
Original source
Further reporting
Where we saw it
Simon Willison
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works