Spawn's 10-person team ships 77,000 monthly code changes via agents

Spawn operates a persistent agent fleet that shipped 77,000 code changes last month, averaging 108 per hour. The company spent $3.1 million on tokens last month, an amount 30 times its payroll, making it one of Anthropic's largest customers. To support this workload, Spawn maintains infrastructure on AWS comprising 5,000 CPUs, 17 terabytes of RAM, and 150 GPUs. The system eliminates traditional pull requests, with agents working directly in the codebase while test suites run every 15 minutes. Over the past year, only about 1,000 of 123,000 changes were reverted.

To manage decision bottlenecks, Spawn deployed AI judges on October 6th. These judges made over 11,000 decisions in eight days, reducing human intervention to 27 calls and scaling fleet output from 1,000 to over 7,000 changes per day. Founder Jacob Sanbury argues that traditional human-shaped workflows hinder AI efficiency, advocating for a model where agents own functions and humans focus on taste and entropy management. Sanbury noted that agents cause fewer production breaks than humans and fix issues quickly when the feedback loop is closed. "Most companies run big human teams that use AI, and we run a small human team alongside a big team of agents," Sanbury said.