I Built the Most AI-Pilled Company in the World

Spawn is a ten-person company building a multiplayer game-creation engine that lets thousands of players change a game in real time. Its founder argues that small teams can gain far more from AI by organizing around agents rather than simply giving AI tools to large human teams. Spawn spent $3.1 million on tokens in one month and says it shipped 77,000 substantial changes over 30 days. Its “fleet” consists of persistent, named agents with memory, permissions and responsibility for areas such as engineering, finance and bug triage; they also launch swarms of temporary helpers. The founder says this approach became viable only with Anthropic’s Fable model, which he considers better than alternatives at maintaining coherence and making sound decisions. Earlier models, he says, either failed at complex work or caused serious product problems.

To keep the fleet productive, Spawn redesigned its engineering practices around agents’ ability to understand the whole system. It eliminated pull requests and lets agents change shared code directly, relying on fast follow-up fixes rather than preventing every imperfect change from landing. Agents diagnose production errors, user reports and requests, while the full test suite runs every 15 minutes; changes are also challenged by helper agents. Irreversible actions, such as payments, still require fixed safeguards. The company reports reverting only 1,055 of 123,548 changes, or about 0.9 percent. It also builds products so agents can observe user activity and replay events, allowing many bugs to be identified and fixed without waiting for a human to report them.