Spawn's Founder Says Permanent AI Agents Shipped 77,000 Changes on $3.1M in Tokens

Disembodied coworkers with names, persistent memory, and all the keys

On his personal site workbyjacob.com, Spawn founder Jacob Sansbury sums up his org chart in a single sentence: half of his employees do not have bodies. Spawn is a San Francisco platform built to let people make multiplayer games with their friends through chat, and Sansbury — who previously founded Pluto before its acquisition by Robinhood, and worked on GeForce Now at NVIDIA and at Bridgewater — runs the studio itself as a live experiment in agentic software engineering.

In his October 29 long-form account, Sansbury describes pairing 10 human staffers with a permanent fleet of AI agents (while his site bio frames the core unit as a six-person company with "three employees, no bodies, all the keys"). Rather than spinning up stateless coding assistants that vanish when a terminal window closes, Spawn gives its resident agents individual names, persistent memory across tasks, and direct permissions inside the production repository.

77,000 changes in 30 days and a token bill at 30x payroll

The numbers Sansbury attaches to that setup invert the usual economics of a seed-stage software shop. Over a 30-day window, Spawn's human-and-agent roster pushed 77,000 changes into the codebase, operating under a rule he summarizes on workbyjacob.com as "every call since the first is in the repo." Instead of optimizing for stingy API usage, the studio poured compute into autonomous iteration.

That velocity came with a self-reported monthly token bill of $3.1 million, a figure Sansbury notes is roughly 30 times the company's human payroll. In a conventional studio, salaries dwarf server bills; at Spawn, human engineers act as high-leverage directors and reviewers for a compute budget thirty times their own compensation.

• Permanent identity and memory

  Agents are treated as standing team members with names and continuous memory rather than disposable chat sessions.

• 77,000 changes in 30 days

  Every agent call and code modification is committed into the repository history, totaling 77,000 changes in a single month.

• $3.1M monthly token spend (30x payroll)

  Compute is budgeted at roughly thirty times human salaries so agents can run continuous build, test, and refactor loops.

What workbyjacob.com reveals about managing a hybrid org

Sansbury has been publishing the operating manual behind this model on workbyjacob.com under the "Thoughts start here" index. Entries from 2024 through 2026 trace the progression from early workflow experiments — such as "Branding for AI Agents: designing a coworker with no face" and "How We Build 2.0: how work moves when models do a real share" — to recent dispatches like "Entities: The New Work (three employees, no bodies, all the keys)", "Phantom Nodes", "Optimizing the Point Away", and "Written for Readers That Didn't Exist Yet".

Those essays highlight a practical takeaway for engineering leaders watching Spawn's experiment: giving agents keys to the repo only works when the repository itself becomes the shared memory log for both humans and machines. When every model call and commit lives in version control, a tiny human crew can audit, steer, and prune tens of thousands of automated changes a month.