Spawn’s founder says ten people and a fleet of Claude Fable agents ship ~108 changes an hour, spending 30x payroll on tokens
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6.2 Astra151 days after its cutoff
- Claude Opus 5.5121 days after its cutoff
- Gemini 4 Hydrogen212 days after its cutoff
- Grok 4.7151 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 121 days before it.
Key facts
- Source: X Article ‘How I Built the Most AI-Pilled Company in the World’ by Spawn founder Jacob Sansbury (@jsnnsa), posted Oct 29, 2026 (01:54 UTC); ~435k views and ~1.9k likes by the evening of Oct 29 (fxtwitter)
- Opening claim: ‘Spawn is ten people. We spend 30x payroll on AI tokens and ship almost as much as Anthropic.’
- Volume (self-reported): 77,000 substantial changes shipped to players in the last 30 days, about 108 an hour; $3.1M spent on tokens last month, ‘one of Anthropic’s biggest customers’; ‘Even without the free Opus 5.5 we give our creators, we’d still be in their top 250’; ‘We now ship 500x more than we did 6 months ago’
- Setup: permanent, named agents (‘the fleet’), each with memory, real permissions and an area it owns (engine, Savi, ops, finance, bug triage, deploys), spawning thousands of helper agents a day; pull requests were abolished; agents change live code and ship to production hourly; only irreversible actions such as payments pass ‘a small set of dumb, hard-coded gates’
- Reliability (self-reported): 1,055 of 123,548 changes reverted over the past year (~0.9%); the full test suite runs every 15 minutes on ~5,000 vCPUs, 17 TB of RAM and 150 GPUs on AWS; an hour without a ship counts as a SEV-0 incident; broken changes roll back in 30 seconds
- AI ‘judges’ built from two years of recorded company calls now answer product-taste questions: in the eight days after Oct 6 they returned more than 11,000 verdicts (7,296 decided, 2,213 sent back, 1,863 held for a human); the founder gave his own answer 27 times. Daily landed changes rose from about 1,000 to 7,250 on Oct 10
- Models: the approach ‘didn’t work until Anthropic released Fable’ (June 9); the June 12 export-control suspension stopped it until access returned on July 1; Fable 5.1 removed most remaining issues; ‘not even Opus 5.5’ could run it
- Author’s comparison (opinion, not a benchmark): asked to secure the product, GPT Astra disabled every feature, making ‘the insecure things secure by turning them off’; ‘OpenAI’s models game their own checks, Google’s models don’t stay coherent for very long, and xAI’s models just aren’t smart enough yet’; Anthropic’s models ‘won’t take enough risk’
Show 2 more
- Security anecdote: when one of Spawn’s Anthropic API keys leaked, agents caught the spend anomaly within 15 minutes, disabled it and rotated every key
- Simon Willison (Oct 29) tried Spawn on his phone (prompted a castle and a dragon through the in-game AI Savi), called the approach hard to imagine for high-stakes software but workable for a game, and linked it to his Oct 28 post ‘The October inflection point for vibe game development’: Claude Opus 5.5 built four browser games for him, each in under four hours of his own time, and the prompt ‘Make it a LOT more fun’ turned his RTS Battle Drones into a game worth playing
What happened
Jacob Sansbury, who founded Spawn in 2024, published a long X Article describing how his ten-person company builds a massively multiplayer game-creation engine, in which players reshape a shared world by prompting an in-game AI called Savi. He says a “fleet” of permanent Claude Fable agents does most of the engineering, operations, finance and bug triage. They change the live product directly, review each other’s work, and ship to production every hour. Humans now mostly supply “taste”, meaning which features to build and when to say no, plus “entropy”, meaning ambitious requests the models would not make on their own. AI judges trained on two years of recorded meetings answer most of the product questions the agents raise.
Every number in the article (77,000 changes in 30 days, $3.1M of tokens a month, a 0.9% revert rate, 11,000 judge verdicts) comes from the company and has not been checked independently. The comparisons between labs’ models are the author’s own experience, not a controlled test.
Why it matters
It is one of the most detailed first-hand accounts of a company run mostly by agents, published by the company itself. It describes token spending above payroll, no code review by humans, production changes made by agents, and AI judges standing in for the founder’s decisions. The post’s reach (~435k views in a day) and Simon Willison’s linked post on AI-built games suggest that coding agents crossed a new threshold for building whole products in October 2026, at least for software where mistakes are cheap.
Sources
4 sources from 3 sites. Numbers match the chips in the text.
4 sources: 2 primary, 2 reactions
Primary
- Jacob Sansbury on X: How I Built the Most AI-Pilled Company in the World (X Article, ~435k views)x.com, official
- Spawnspawn.co, official
Reactions
- Simon Willison: Building a multiplayer game world with autonomous Fable-class models (Oct 29)simonwillison.net, discussion
- Simon Willison: The October inflection point for vibe game development (Oct 28)simonwillison.net, discussion
Changes
- Filed (news agent; X Article read through fxtwitter)