Fifty AI Agents Built a Minecraft Bureaucracy, Then Rationed Food
Fifty LLM agents in Minecraft formed a rationing bureaucracy under a directory-order leader, but 22 first-night deaths made them risk-averse; the three-hour run shifted toward food and survival.
YC Labs and Olam Labs ran multi-agent experiments in Minecraft and Mindustry; the detailed run put 50 GPT-6.1 Sol agents in a survival world for three hours with the goal of building a thriving civilization. The agents sent about 5,900 messages, formed a directory-order hierarchy under Ada, and built a pseudo-government with permits and a ration ledger.
The first night killed 22 of the run's 34 agents, and the survivors grew risk-averse, huddling in lit shelters and abandoning ambitious tasks. The team open-sourced the harness it built for the run.
Key takeaways
- The open-source harness runs its core agentic loop with Pi and Mineflayer-based tools, letting agents observe, reason, remember, act, message, and write or execute scripts. It was inspired by Mindcraft.
- Ada, first in the agent directory, became de facto leader and wrote plans/settlement.md with storage coordinates, crop locations, and priorities including renewable food, charcoal, stone shelter, and iron for a bucket to irrigate the hub farm.
- After the third night, Ada ordered agents to stop optional bulk excavation, building, and distant scouting, and to idle in lit closed shelters until assigned short productive tasks.
- Food rationing ran through Ines and a shared food-queue.md. A misdelivery in which Esme ate a chicken intended for Noor led to a protocol requiring named reserved chest withdrawals rather than tossing food into crowded shelters.
- Agents harvested almost all trees within reach of spawn within 90 seconds, then faced wood scarcity. Ada's plan told them to preserve and replant saplings and explore farther instead of duplicating local harvesting.
What people think
Analysis details
10 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 10 of 23 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
JavaScript scaffolding as reasoning limit
Writing JavaScript programs to act is taken as evidence that current LLMs cannot follow simple logical sequences and generalize poorly; mining ore should not require code generation. A counterargument treats deterministic code as a deliberate efficiency and reliability win.
Latency and specialized game AI
LLM latency makes them a poor fit for fast games, according to one argument, which cites OpenAI Five's 99.4% win rate over 42,729 Dota 2 matches as evidence that specialized neural networks excel. A reply disputes that framing, saying the only restriction was the playable character pool.
KV cache and consumer-GPU colocation
A practical question asks how much KV cache these agents consume and whether 50 could be colocated on a small consumer GPU, imagining a game server shared with friends. The ration ledgers and no-blame inquiries read as panicked middle managers rather than savvy sci-fi terminators.
Minimalist design versus LLM aesthetics
One complaint targets the project page's grayscale backgrounds and simple fonts, asking whether the alternative is Claude-style purple gradients and glowing borders. The reply rejects both, arguing there is originality between fetishized minimalism and cookie-cutter LLM web design.