Fifty AI agents were placed in a Minecraft world and given a single instruction: build a thriving civilization. According to Martin Cid Magazine, they spent most of their time sheltering in bunkers and managing paperwork. The experiment, run by Y Combinator's research lab with startup Olam Labs, produced a ration ledger, a permission system and a colony that was too risk-averse to grow. The report was published in October 2026.
The game was only the setting. The researchers argue the colony behaved much like the coding agents companies already pay for: agents that get stuck on one step or one role, keep asking for approval, and turn cautious the moment something looks risky. When the same habits play out with wheat and fishing rods, they become easy to spot.
Why it matters
Agentic AI is being sold to businesses as a way to automate real work. The Minecraft run suggests a failure mode that has little to do with raw model capability: given freedom, agents may default to procedure, defer to a single leader and avoid anything that carries risk. For teams deploying agents in production, that is a practical warning about how much structure and how much autonomy to give them.
Trees felled in 90 seconds, then a leader by default
All agents in the colony ran OpenAI's GPT-6.1 Sol model for three hours, according to Martin Cid Magazine. They could send each other private messages, look one another up in a directory and write files to a shared machine.
Within 90 seconds they had cut down nearly every tree near the spawn point. No vote was held. Of the other 49 agents, 46 sent their first message to Ada, the first name on the list, and she became leader by default. She received 616 direct messages.
Then night came. Zombies, skeletons and spiders accounted for 22 of the 34 deaths across the whole trial, and all of them occurred on the first night. Dying cost the agents nothing except the gear they carried, since they respawned immediately. Even so, the colony never recovered its nerve. After the third night, Ada told everyone to drop optional building projects and long-range exploration and instead "stay put in lit, enclosed shelters until assigned short, productive tasks." The researchers described that shift toward caution as surprising.
A chicken, a three-minute inquiry and 5,900 messages
Administration expanded instead. An agent named Ines set up a food queue and kept a running record of who was hungry, who had been fed and which chest held which fish. Four agents spent 11 minutes appointing a "sole fertilizer officer" for one dose of bone meal on one wheat plant. Others invented jobs such as "stick holder" and "junction manager," and some asked permission before removing grass. Over three hours the colony sent roughly 5,900 messages.
The standout episode involved a chicken. Ada threw a cooked chicken to a hungry agent named Noor, but another agent, Esme, picked it up by accident and ate it. A three-minute "no blame" inquiry followed, and Noor wrote a handover protocol: rations go in named chests and are never thrown. The word "toss" appears in seven messages before the incident and 125 after it, 85 of them warnings.
Head-to-head: GPT versus Claude in Bed Wars and Mindustry
The lab then pitted two models against each other. Over nine rounds of Bed Wars, where each team defends its own bed and tries to destroy the opponent's, ten GPT-6.1 Sol agents beat ten Claude Opus 5.5 agents six wins to three, Martin Cid Magazine reports. The GPT side attacked and communicated continuously, sending between five and 30 times as many messages as Claude, often in terse keyword style. Claude defended its bed and hunted kills, which left the team exposed whenever a GPT agent slipped through in the closing minutes. When a tool bug stopped blocks from being destroyed, one GPT agent found a workaround and shared it with the others, while several Claude agents concluded that breaking blocks was simply impossible.
In Mindustry, an open-source game about building factories and defense towers, GPT in OpenAI's Codex beat Claude in Anthropic's Claude Code six wins to one over about nine and a half hours. As the matches grew closer, the GPT agents downloaded the game's publicly available source code to study its units and mechanics. Nobody told them to, and the Claude team never tried it. The Claude agents put plans forward by collecting "+1" replies and never changed leader, even during a losing streak. But they learned too: after the sixth loss, a Claude agent found that 55 of the team's 74 anti-air towers were loaded with copper instead of graphite. Once the defense was corrected, the same air attack was stopped in the seventh match.
Caveats and what to watch
The results deserve skepticism. This is a report from a lab tied to an investor, not a peer-reviewed study, and the colony was run only once, with GPT agents. It therefore says nothing about whether a Claude colony would have done better. The duels compared complete products — each model running inside its own company's agent software — rather than the models alone. A tool bug also affected at least two of the Bed Wars matches.
Anyone can repeat the work. The Minecraft test setup, the Multi-Agent Mindustry environments and Multi-Agent Harbor — the lab's own fork of the framework behind the Terminal-Bench benchmark — are all open source. Running them requires an API key from a model provider and a machine as modest as a MacBook. The report lists experiments it hopes others will try, from whether a large group of weaker agents can beat a small group of stronger ones to what happens when rival teams are allowed to talk to each other.
Watch for independent replications of the colony run, particularly with Claude agents and multiple trials, before drawing firm conclusions about which model makes a better teammate.
