The Agent case that raised my eyebrows a dozen times: no PRs, 108 changes per hour

Simon Willison shared a long-form article by jacob, calling it the most cutting-edge case of agentic engineering in the Fable era. He said his eyebrows raised at least a dozen times while reading it. The original link is at the end.

From the thread, Spawn is a game platform with a chatbot called Savi in the game. Users can use prompts to generate things like castles and dragons into the game. Simon tried Spawn on his phone: first he prompted a castle and a dragon, then rode the dragon around the castle.

Luis Catacora made the article into a 2-minute explainer video.

What really excited the comments wasn't that AI can write game code. It was how far the process had been changed.

One reply mentioned that the team let permanent agents have real permissions, and there are no PRs. Someone followed up asking exactly what scope each agent can touch and how permissions are divided. Another reply called Astra ensuring security by turning off all features the funniest part of the whole thing: no features, no security problems.

Another point discussed repeatedly was AI judges making taste judgments. Louie Mota summarized that two years of call records had been turned into taste; of 7,296 call decisions, jacob personally intervened only 27 times. He added a biting line: most people don't even write down their own decisions, so they have no material to build a judge.

This is more worth looking at than model parameters. The ceiling of agent engineering depends on how many replayable traces of decisions you leave behind. Without records, there is no learnable judgment, and in the end you can only have agents do more mechanical work.

The scale numbers are also striking. Someone in the replies mentioned 108 changes an hour, with a team of only ten people. Someone else asked, with 108 changes an hour, who is actually reading them, and has review also been handed off to a bot? Another detail was 700 issues waiting overnight; one reply joked that this is like every sales manager's Monday, just with better grammar.

Some people got stuck on the business side. Caleb LeNoir said Spawn reaching over $32 million in annual revenue was something he couldn't figure out. How could a game he hadn't heard of until that day support these token costs? orca apologist asked the same question. The thread didn't answer this question, but it reminds people that no matter how sharp the case looks, someone still has to pay the bills.

There is also a passage in Jacob's reply that says a lot about the team's state. Nikita Bier said mobile resolution wasn't solid enough, and jacob responded that a lot of Fable tokens would be spent on this, with the goal by December of achieving GTA 7, 1-second loading, 60fps, and native resolution on crappy Android devices—provided Anthropic doesn't keep F5.5 stuck. He had also said earlier that he wanted to spend $1 million on Fable 5.5 tokens to make Spawn games faster.

These details make the case more than a demo. It is demonstrating an organizational model: agents entering permissions, reviews, taste judgments, and operations issue queues. The most important question isn't whether to copy the setup, but how to define each agent's scope, how to audit, and who is responsible when something goes wrong.

One reply said that after reading this article, going back to copying error messages into a chat box feels like returning to the Stone Age. This self-deprecation is pretty accurate. The gap isn't in the tool names; it's in whether the team has turned decisions, permissions, and feedback into something agents can pick up and run with.

Original: https://x.com/simonw/status/2115821668350435645

jacob's long-form article: https://x.com/i/article/2115540320373104640