A village of 30 villagers who live by two kinds of decisions. When something happens to a villager, the Jev model picks their next action on the spot, with odds (fast thinking). About once a real hour a Claude council sets goals and village rules, and looks back on events and impulses to add or remove a villager’s acquired traits (slow thinking). Nothing grows in winter, every choice is shown with its odds, and a viewer can whisper to one villager once a day.
What AI handles
- Picks a villager’s next action on the spot, with odds (the Jev model). About once an hour, sets the village’s goals and rules and adds or removes villagers’ traits after looking back on events (the Claude council).
What a person decides
- Which actions are on offer and what the AI is told, fixed in code. Steal and attack are available only under hunger, a grudge, or a rumor, and each ration level is passed with words like “enough” or “not enough”.
Figures
- Share of steal actions (first run → after limiting options in code)as of 2026-09
- 73% → 5%
- Time to decide one tick (all villagers in parallel, median)as of 2026-09
- 620 ms
- Failed callsas of 2026-09
- 3 timeouts in about 1,500 calls
- A starving villager’s chance of stealing (no added trait → “bitter” and “distrustful of everyone”)as of 2026-09
- 17% → 51%
- Villagers who starved with food still in the granary (first headless run)as of 2026-09
- 16 over two winters
- Estimated tokens per dayas of 2026-09
- About 83M with one viewer, about 10M with nobody watching
Hypothesis
If slow thinking looks back on fast impulses and rewrites a villager’s character, and that character shapes the next impulses, AI villagers change visibly over the short and long term, and people come back to watch the change.
Learnings
- Each villager's choice fit its situation, yet 73% of all actions were steal and the village starved in about 110 ticks. The first fix wrote each action's consequence into the criteria and added a food-outlook key to the state, and steal rose to 80%.
- Offering steal and attack only under hunger, a grudge, or a rumor, decided in code, cut steal to 5%. In a 200-tick run a famine took the village from 30 to 15, and then it recovered.
- Deciding for every living villager in one tick, with the calls run in parallel, took a median of 620 ms of wall time. Only 3 of about 1,500 calls timed out.
- One continuous viewer would use about 83M tokens a day, and an empty village still about 10M, so the village now stops while nobody is watching.
- Holding the situation fixed and changing only the traits the council adds moved Jev in the direction of the trait. A starving villager’s chance of stealing was 17% with no added trait and 51% with “bitter” and “distrustful of everyone”, and mood fell from 2.3 to 1.2. “Caring” cut stealing to 7% and doubled farm work; “resigned” tripled prayer.
- Jev leaned hard on the plan it was shown. Even when starving it rated “carry on with the plan” around 0.6, so a rule of always keeping the plan above 0.5 left a villager leaning toward theft still chatting on schedule. Making that rating the probability of keeping the plan raised thefts in the spring famine from 1 to 7.
- In the first headless run the council locked the granary and rationed two meals a day, and 16 villagers starved over two winters with more than 800 food still stored. The council had never been told how many meals a day are enough; once each ration level came with words like “not enough” or “enough”, it raised rations to four a day after the first deaths.
- The council tended to read experience charitably. In a six-hour headless run, 15 of 28 trait changes were “after being helped”, and the village mostly grew more cooperative. The famine came in the spring and summer after the first winter; once 30 had fallen to 14, nobody died for more than two years.
Stack
Cloudflare Workers / Durable Objects / Workers AI (TypeSafe Jev) / Claude API / React / TypeScript