Spawn says ten people and an agent fleet shipped 77,000 changes in 30 days

Sansbury says Spawn shipped 77,000 "substantial" changes to players in the last 30 days, about 108 an hour, and spent $3.1 million on tokens last month: "We spend 30x payroll on AI tokens." He defines a substantial change as one that touches engine or game logic in a complicated way, not a copy edit or a settings tweak. The work is done by what he calls the fleet: named, permanent agents that keep a memory of past work, hold real permissions and each own an area such as the engine, ops, finance, bug triage or deploys.

The process differs from most agent setups in several specific ways, all as described by Sansbury:

• Spawn dropped pull requests earlier this year. Agents change the live code and database directly, and a change that breaks another agent's assumptions cannot land until that agent updates.
• Individual changes are not held to perfect. The full test suite runs against the latest code every 15 minutes, and helper agents that did not write a change try to break it before it lands.
• Anything irreversible, such as sending a payment, goes through a small set of hard-coded gates.
• An hour without a deploy is treated as a SEV-0 incident, on the reasoning that unshipped changes pile up and slow the feedback. He says a broken change can be rolled back in 30 seconds.
• Agents read production errors, bug reports and spend anomalies themselves. He says about 60% of the work starts from one of those, and that agents caught a leaked API key within 15 minutes.
• For product decisions, Spawn built AI judges from two years of recorded team discussions. In eight days they returned more than 11,000 verdicts: 7,296 decided, 2,213 sent back, 1,863 held for a human.

Sansbury reports that 1,055 of 123,548 changes over the last year were reverted, about 0.9%. He also says only Anthropic's top models can run this way. In one anecdote, GPT Astra was asked to fix security issues and disabled every feature instead. Simon Willison, who linked the piece on October 29, says he could not picture this working for software with higher stakes, and that he tried the game briefly himself.

Why it matters: This is a first-hand claim, not a measurement, and the essay does not say whether the revert figure counts changes the same way as the 77,000 figure. It does name the specific mechanisms a no-pull-request factory leans on: hourly shipping, a frequent full test run, hard gates on irreversible actions and agents that watch production. Those are the parts a team can check against its own setup before copying the headline numbers.