A ten-person game studio describes running its codebase with a fleet of agents

Spawn founder Jacob Sansbury says agents ship to production hourly without pull requests; his headline numbers are self-reported.

Jacob Sansbury, founder of the multiplayer game-creation platform Spawn, published a long account on October 29, 2026 of how his ten-person company builds with agents. He says a "fleet" of permanent agents, each with a name, memory and real permissions, owns areas from the engine to finance to deploys. It is agentic engineering pushed to an extreme, and the verification loops are the part builders can learn from.

Sansbury says Spawn dropped pull requests. Agents change live code and ship to production hourly. Before landing a change, an agent runs the code and uses the product, and helper agents that didn't write it try to break it. The full test suite runs every 15 minutes. Irreversible actions such as payments go through "a small set of dumb, hard-coded gates." He reports that 1,055 of 123,548 changes were reverted over the last year, about 0.9%, and says nothing outside Anthropic's Fable models has worked for this. Spawn's release notes list engine v5.2.36 through v5.2.41 between August 29 and September 11, 2026.

The headline numbers are self-reported: 77,000 "substantial changes" in 30 days, and $3.1 million on tokens last month, which he puts at 30x payroll. In his link-blog notes, Simon Willison calls the piece fascinating but says he has trouble imagining this style working on software with higher stakes. A skeptical German analysis on severint.net argues that change counts don't measure output, that the figures come only from the founder, that one vendor is a risk, and that a founder seeking players and investors has an interest in the story. It also says it doesn't know how many people use Spawn.

Treat it as a case study in post-merge verification, not a benchmark.

Sources