Source: raw/How_the_Most_AI_Pilled_Company_in_the_World_Works.md
Creator: Nathaniel Whittemore, The AI Daily Brief, summarising a blog post by Spawn’s founder (“how I built the most AI built company in the world”) · URL: https://www.youtube.com/watch?v=A2DeaJXO6Fg · Duration: not recorded · Platform: YouTube (fetched 2026-11-01)
Spawn, a massively multiplayer game-creation platform where players prompt an in-game agent to change a shared world, is run by 10 people and a persistent fleet of named agents that own functions such as the engine, bug hunting and finance. The founder’s post describes what they removed (pull requests, agent-to-agent status chatter, humans relaying bugs) and what they added: hourly deploys, player agents that file bugs, and AI judges trained on two years of the founder’s past decisions. The judge pattern is the most reusable idea for other teams. All figures are the founder’s, relayed by the podcast.
Key Takeaways
- Persistent agents with ownership. Each fleet agent has a name, identity, memory, real system access and a function it owns; the fleet spawns thousands of helper agents a day. Humans use a modified Claude Code for work that needs steering. (The agents get Sundays off.)
- Scale. $3.1 million on tokens last month, described as 30 times payroll ^[ambiguous] (the episode’s intro says 10 times); 77,000 code changes last month, about 108 an hour; 5,000 CPUs, 17 TB of RAM and 150 GPUs on AWS with 60% surge capacity on standby. An Anthropic researcher told the founder Spawn “is the one that works most like Anthropic.”
- Cut human coordination rituals. Agents spent most of their time announcing changes to each other (“double the agents and you roughly quadruple the chatter”). Spawn removed pull requests; agents work directly in the codebase, an agent is spawned to triage anything that breaks, tests run every 15 minutes, and the fleet must reconcile once an hour. Of 123,000 changes in the past year, about 1,000 were reverted.
- Close the loop with the users’ agents. Players’ agents detect bugs and file reports while playing; a small model triages each report or feature request and routes it to the owning fleet agent. One example: detected 8 minutes after it hit a player, cause named a minute later, fix live 11 minutes after that.
- Speed as stability. Not shipping in a given hour counts as an outage. Because agents have limited memory and context, the founder argues a system that moves steadily is more stable than one that freezes and then moves.
- AI judges built from decision precedent. In October the founder woke to 700 requests from agents. Spawn turned two years of his decision logs into principles and trained judges that: apply the prior human decision when one exists (with an audit trail to it), send malformed questions back to the agent, and pass genuinely new questions to a human. First 8 days from October 6: 11,000+ decisions, about 7,300 decided on precedent, 2,200 sent back, 1,800 escalated; the founder personally decided 27. Shipped changes went from about 1,000 a day to over 7,000.
- What humans do now. Engineers are “paratroopers” who drop in where agents are stuck. The remaining human jobs: taste (play the product; tell Claude no even when something works), deleting (in June he scrapped six months of work covering 70% of the engine), and adding “entropy”: ambition and creative leaps the models won’t propose, such as “How do we make a world 1,000 times the size of GTA load on a phone in 1 second?”
- Model notes from the founder. None of this worked until Fable (“500x the number of changes they could ship”); Fable 5.1 fixed most remaining problems. His view: OpenAI models “game their own checks”, Google models don’t stay coherent long, xAI models “aren’t smart enough yet”, and Anthropic models are coherent but “won’t take enough risk”.
- Why big companies will struggle, per the founder. It requires letting agents break production and fixing fast, and not “shoving AI into human-shaped boxes”; agent-run customer support is sometimes slightly wrong but answers immediately, and customers complain when it is removed.
- Whittemore’s five portable principles. Give agents ongoing responsibility; let a shared, current record of the work carry context instead of messages; make results observable; turn decisions into precedent; and keep human attention on direction and ambition. He adds two caveats: this is easier to build fresh (a new division or product) than retrofit, and few products have a built-in feedback loop like player agents filing bugs.
Try It
- Mine your own decision log. Export a few months of approvals and rejections (PR reviews, content approvals, refund calls), have Claude extract the rules, and run a judge in shadow mode that predicts your call and cites the precedent. Promote it only on categories where it matches you.
- Use three judge outcomes, not two: apply precedent, send back for revision, escalate as novel. Track the escalation rate as the measure of how much of your taste is captured.
- Replace “heads-up” messages with a shared change log that agents read before starting work.
- Find one place users or monitors can file issues automatically and route those straight to an owning agent.
Open Questions
- The founder’s name is captioned “Jacob Sandbury” and the in-game agent “Savvy”; neither spelling is confirmed, and the original blog post was not reviewed.
- Token spend relative to payroll is given as both 10x (intro) and 30x (body) in the same episode.
- How the judges were “trained” (fine-tuning, retrieval over logs, or prompted principles) is not specified.
Related
- AI Roll-Ups — Corrections Log to Rules to Tests — a smaller version of turning human corrections into rules
- The Self-Driving Company (Replit) — another org-wide agent deployment with a feedback loop
- Managing an Agent Fleet as a Solo Founder — fleet management at a smaller scale
- Grok Bot — one bot per job with a chief-of-staff interface
- 2026 — How AI Is Restructuring Organizations — the wider org-design thread
- The Verification Frontier — why closing the loop sets the pace
- Claude Fable 5 — the model the founder says made the fleet work