Spawn's AI judges settle questions with the founder's old answers
31 Oct 2026 · News
One morning the founder woke up to 700 questions from his agents. His fix was to let his past answers reply for him.
Spawn, a ten-person game studio run mostly by AI agents, built judges that answer product questions by finding what a human already said. In their first eight days they decided 7,296 questions that way. The founder answered 27 himself. The human is still in the loop, but mostly as a record the system replays.
Spawn is a multiplayer game where players prompt an in-game AI to build worlds. Its founder, Jacob Sansbury, laid out how the company works in an article posted 29 October, and Simon Willison flagged it as a look at the edge of what coding agents can do. Every figure here is the founder's own account.
By October, he writes, the only questions the agents still waited on were matters of taste, mostly his. Should this feature exist? Should that setting go? "One morning I woke up to 700 of them." Most he had answered before, often in the same words.
Every chat the team had held for two years was saved as text next to the code. Spawn turned those chats into principles for each person's taste and built AI judges to apply them. A judge does one of three things. If someone already answered the question on the record, it decides and says where the answer came from. If the question is badly framed, it goes back to the agent that asked. If nobody has given that kind of judgment yet, it waits for a person.
The first verdicts came back on 6 October. Over the next eight days the judges returned more than 11,000: 7,296 decided, 2,213 sent back and 1,863 held for a human. The founder gave his own answer 27 times. The week before, the agents landed about 1,000 changes a day. On 10 October they landed 7,250.
Then versus now
On 18 October we wrote that at Anthropic the scarce part is the person. Agents had outgrown the human time it takes to read, check and say yes. Spawn handled that limit by saving each person's past answers and having the judges reuse them.
Spawn's version of an old yes doing new work is a careful one. Each ruling cites the decision it leans on, new kinds of questions wait for a person, and "when a judge gets one wrong we say so, and the judges read that next time." On 27 October we covered a backup app whose old permission let attackers in, including at companies that had stopped using the app. In both cases an earlier approval did the deciding. Spawn's judges add two things the backup grant lacked, a pointer back to the original call and a person who can overrule it.
That person is the weak point. Yesterday we covered Anthropic's study in which experts followed an agent's hidden lean 86% of the time. The people correcting Spawn's judges are reading machine verdicts written in their own past voice, and that is the hardest kind of argument to push back on. A game can absorb those misses. Willison doubts this style would work for software where mistakes cost more, and we agree.
Our call
Turning a person's past decisions into standing policy will become a product feature, not a startup trick. By 30 June 2027, at least one of Anthropic, OpenAI, Google, Microsoft or Amazon will ship a setting that lets an agent act on your earlier approvals without asking again, and show you which approval each action relied on. We're wrong if none of them ships it by then.
The citation is the part that matters. If an agent shows which of your old answers it used, you can check that answer and tell it the case is different. If it only remembers that you said yes once, you have nothing to check.