- Y Combinator Labs and Olam Labs put 50 GPT-6.1 Sol agents into Minecraft for 3 hours: after the first night (22 deaths) the colony moved from building to rationing food.
- In team matches GPT-6.1 Sol beat Claude Opus 5.5 6–3 in Bed Wars and 6–1 in Mindustry.
In short: in October 2026 Y Combinator Labs and Olam Labs published a report called “Agents in Games”. They put large groups of AI agents into games and watched what they did together. The result reads like satire about corporate life, but the lessons are serious.
Civ50: 50 agents, one world
- 50 GPT-6.1 Sol agents got 3 hours in survival Minecraft and the instruction “Build a thriving civilization”. Each AI agent had a directory of the other players, a shared sandbox and direct messages.
- They exchanged about 5,900 messages. 34 agents died, 22 of them during the first night.
- Ada became the leader simply because she was first on the roster. There was no vote: 46 of the other 49 agents sent their first message to her, and she received 616 messages in total.
- After the first night the colony dropped ambitious building in favour of food: a ration queue, a delivery log and fishing shifts.
- When a tossed chicken was picked up by the wrong agent, the word “toss” went from 7 messages to 125, and the group wrote a food handoff protocol. Four agents spent 11 minutes appointing a sole operator for a single piece of bone meal.
Teams: GPT versus Claude
- Bed Wars: 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents over nine 45-minute games. GPT won 6–3 and sent 5–30 times more messages.
- Mindustry: GPT-6.1 Sol in Codex beat Claude Opus 5.5 in Claude Code 6–1 over about 9.5 hours. In game 6 Claude loaded 55 of its 74 anti-air turrets with the wrong ammunition; in game 7 it built a ring of 207 correctly loaded turrets and stopped the same attack.
How to read this
This is a multi-agent system in a game, not in a company, and a single experiment without repeats. The authors built their own harness and released it as open source. They themselves see a parallel with coding agents, which can also get stuck on one step and become overly cautious, especially around security and production changes. I write about agents in the Coding agents hub.

