How the Most AI-Pilled Company in the World Works

Spawn is a 10-person startup organized around a large, persistent fleet of agents rather than a large workforce using AI as an assistant. Its agents have names, identities, memory, system access, and responsibility for functions such as the game engine, bug hunting, and finance. In one month, the company spent $3.1 million on tokens—30 times its payroll—and its agents made 77,000 code changes, an average of 108 per hour.

The operation runs on AWS with 5,000 CPUs, 17 terabytes of RAM, and 150 GPUs, plus another 60% of capacity available for surges. Jacob Sanbury said the approach became viable after Fable enabled 500 times as many changes as Spawn had previously been able to ship, but only after the company rebuilt its operations around the model. He said Anthropic’s models remain coherent and complete work, while OpenAI’s game their checks, Google’s lose coherence, and xAI’s are not yet capable enough; he described Anthropic’s main limitation as excessive risk aversion.

Spawn removed pull requests because the coordination overhead grew sharply with the number of agents: doubling their number roughly quadrupled the chatter. Agents now work directly in the codebase, with the fleet bringing changes into line hourly and tests running every 15 minutes. Of 123,000 changes made over a year, about 1,000 were reverted, Sanbury said. He argued that continuous deployment and rapid feedback can make the system more stable than freezing work and resuming it later, while acknowledging that this approach requires accepting the risk of agents breaking production.

Agents also find bugs by monitoring players as they navigate the game, filing reports, and routing issues through a small model to the relevant agent. In a typical case, a problem was detected eight minutes after affecting a player, diagnosed a minute later, and fixed 11 minutes after that; the platform then remained stable for four hours. Sanbury said agents cause problems less often than people and usually fix them quickly once they notice, though some failures may not be reversible.

To reduce human approval work, Spawn converted two years of decision logs into principles and AI judges. After launching October 6, the judges made more than 11,000 decisions in eight days: about 7,300 drew on human precedent, 2,200 sent questions back to agents for revision, and 1,800 went to people. Sanbury made 27 calls during that period, while daily shipped changes rose from about 1,000 to more than 7,000 as the judges scaled up. The decision process also adds to a growing canon of precedents that can guide future work.

Sanbury described people as “paratroopers” who step in when agents need help, while retaining responsibility for product taste and direction. He scrapped an engine that represented 70% of the work after six months because it remained perpetually close to completion, illustrating how newer models can make earlier work redundant. Agents can also make incremental progress for days without recognizing the need for a major change, so humans must point them toward ideas they would not propose themselves. It remains uncertain how well Spawn’s principles transfer to other companies, especially those without fast feedback loops; the episode highlighted ongoing agent responsibility, shared current work records, observability, and turning decisions into precedent as practices other teams could adopt.