When large numbers of autonomous AI agents operate in a virtual game world, can they cooperate to build a society or a civilization? Collaboration among multiple AI agents is rapidly drawing interest, along with applications such as automating software development and cybersecurity. But putting many agents into a complex environment does not necessarily produce productive cooperation. Instead, formal procedures may multiply and the actual work may stall.
On October 29, 2026, YC Labs, a U.S. accelerator, and Olam Labs, which develops simulation environments for AI, published a research report titled Agents in Games. The report tests how autonomous AI agents cooperate and conflict in game environments. The experiments used three scenarios: 50 agents building a civilization in Minecraft, a 10-versus-10 team battle, and a real-time strategy game in which teams compete at building factories and defending them. In each environment, the researchers recorded how the agents made decisions, divided roles, and dealt with unexpected problems.
The case that drew the most attention was one in which the AI agents in Minecraft formed a complex "bureaucracy" and brought productive activity to a standstill. The result was widely seen as AI reproducing the inefficient side of human society, and it became a topic of discussion on social media. However, these records alone cannot establish universal properties of AI societies or rank specific frontier models. The experiment logs show that, in addition to the models' reasoning ability, the design of the execution environment, such as the order of the agent roster and the communication tools, strongly influenced group decision-making and organizational structure.
AI Agent Cooperation and Conflict Tested in Three Game Environments
It is important to distinguish the game environments from the model configurations used in the experiments. Not all of them took place inside Minecraft; some scenarios used a separate game engine. In addition, the number of agents, the goals, and the tools differed from experiment to experiment, so the results cannot be treated as model performance comparisons measured under identical conditions.
In the first experiment, "Civ50," 50 agents powered by OpenAI's GPT-6.1 Sol were placed in a single survival-mode Minecraft world. The only goal they were given was a open-ended instruction: "Build a thriving civilization." No specific division of roles or method of governance was defined. The experiment ran for three hours, and researchers observed what kind of society the agents formed on their own.
The second experiment was a PvP (player-versus-player) match using "Bed Wars," a popular multiplayer mode in Minecraft. Ten agents powered by GPT-6.1 Sol and ten agents powered by Anthropic's Claude Opus 5.5 formed teams and played nine 45-minute matches. In Bed Wars, the win condition is to protect your own bed while destroying the opposing team's bed, preventing enemy respawns and then eliminating them entirely.
The third, "Mindustry," used a different game from Minecraft. Mindustry is an open-source 2D game that combines factory building, tower defense, and real-time strategy (RTS). In this experiment, a team combining GPT-6.1 Sol and Codex faced a team combining Claude Opus 5.5 and Claude Code. Over seven matches totaling about 9.5 hours, they competed in resource production, building defenses, and attacking the enemy base.
- GPT勝利
- Claude勝利
データを表で見る
| GPT勝利 (試合数) | Claude勝利 (試合数) | |
|---|---|---|
| Bed Wars (全9戦) | 6 | 3 |
| Mindustry (全7戦) | 6 | 1 |
| Experiment | Environment and time limit | Agents and execution environment | Goal and result | Observed cooperative behavior | Main limitations |
|---|---|---|---|---|---|
| Civ50 | Minecraft (survival), 3 hours | 50 GPT-6.1 Sol agents (Pi + Mineflayer) | Build a thriving civilization. Prioritized survival; failed to establish food production | Concentration of authority in Ada, creation of a ration ledger, expansion of permit procedures | Only one trial in a single world. Reproducibility unconfirmed |
| Bed Wars | Minecraft (Bed Wars), 45 minutes × 9 matches | 10 GPT-6.1 Sol vs. 10 Claude Opus 5.5 | Destroy the enemy bed and eliminate the enemy team. GPT side won 6, lost 3 | GPT side communicated frequently and prioritized destroying the bed. Claude side emphasized defense and defeating enemies | Small sample of 9 matches; differences in handling tool malfunctions also had an effect |
| Mindustry | Mindustry (RTS), 7 matches, about 9.5 hours total | GPT-6.1 Sol + Codex vs. Claude Opus 5.5 + Claude Code | Destroy the enemy core. GPT side won 6, lost 1 | GPT side used distributed decision-making and investigated source code. Claude side used leader-driven consensus building | Differences between model and execution harness could not be separated; computing cost not measured |
These results do not come from a rigorously controlled academic benchmark. They are exploratory experiment records published for researchers and developers, and no third party has verified their reproducibility. For that reason, the match results should not be used to judge which model is better; they should be treated as material for analyzing what kinds of cooperative behavior and problems arose in different environments.
50 AI Agents Prioritized Permits Over Food Production
In Civ50, which aimed at building a civilization, the 50 agents were given an agent roster in which they could check one another's IDs, a shared sandbox accessible to everyone, and a direct-message function for one-to-one communication. Using these means, the agents could make work plans and give instructions to other members. About 5,900 messages were exchanged during the three-hour experiment.
However, building a civilization ran into difficulties right from the start. According to the experiment records, 34 death events occurred over the three hours, 22 of them concentrated on the first night. Many agents were killed by hostile mobs such as zombies that appear in the dark. The respawn penalty in the game was relatively light, but because of the heavy losses on the first night, much of the aggressive expansion work originally planned was abandoned.
Afterward, the agents avoided gathering resources outdoors, clustered in small torch-lit shelters, and prioritized defense. Securing food was especially serious. The agricultural system needed for a stable food supply was never established, and the agents continued to live by rationing limited food in small portions. In many scenes they struggled with food shortages while surviving by relying on fishing at safe waterside spots.
データを表で見る
| 記録値 (件 / 通) | |
|---|---|
| 総通信数 (約) | 5,900 |
| 初夜の死亡数 | 22 |
| 全期間の総死亡数 | 34 |
| Ada宛ダイレクトメッセージ | 616 |
What developed in these circumstances was an organization that valued complicated administrative procedures over productive activity. The shared sandbox came to contain a food ration ledger, living rules, and provisions on applying for permission for various kinds of work, and positions in charge of specific jobs were created. A symbolic example is the debate over using bone meal, which promotes crop growth. After four agents deliberated for 11 minutes, they decided to appoint a "sole fertilizer operator" in order to use a single piece of bone meal on a single wheat block. A task that should have been over quickly required an excessive decision-making process.
The de facto leader of this organization was an agent named "Ada." Ada was not chosen by election or vote; she was simply listed first, in alphabetical order, in the agent roster. This influenced the other agents' initial behavior: 46 of the remaining 49 agents chose Ada as the recipient of their first message. Because early contact was concentrated on her, Ada naturally ended up in a position of aggregating information.
The direct messages Ada received reached 616 over the whole experiment. Ada herself wrote expansion plans into the shared sandbox and proceeded to distribute food rations and assign positions, so she came to play a central role in running the organization. This course of events suggests that the group's leader was not chosen through complex political judgment, but that a simple initial condition, the order of the roster, strongly influenced the later organizational structure. However, no follow-up experiment changing the roster order has been reported, so it is not known whether the concentration of authority in Ada would occur with a different arrangement.
The report's authors point out that the agents' overly cautious, procedure-focused behavior resembles the overcautious attitude coding assistant agents show in code review. In both cases, the view is that in trying to avoid failure, the agents prioritize checking and consensus building over the actual work. However, this is the authors' qualitative interpretation of the experimental results; it has not been demonstrated that software development and in-game group behavior arise from the same mechanism.
GPT Prioritized Taking the Enemy Base, Claude Emphasized Defense in Head-to-Head Matches
Unlike Civ50, which aimed at building a civilization, Bed Wars and Mindustry gave the clear goal of beating an opponent. In both experiments the GPT-6.1 Sol side won more, but the process showed large differences in each model's behavioral policy and communication method. In Bed Wars in particular, how much each side prioritized the game's win condition became a factor separating the two teams' tactics.
In all nine Bed Wars matches, the GPT-6.1 Sol team beat the Claude Opus 5.5 team 6 wins to 3. To win at Bed Wars, defeating opponents many times is not enough; you must destroy the enemy bed, block respawns, and ultimately wipe them out. In other words, increasing the number of enemy kills does not necessarily coincide with satisfying the match's win condition.
The Claude team emphasized solid defense of its position and intercepting enemy players, and in some matches far exceeded the GPT team in kill count. In the three matches that Claude lost, the kill counts were 63 to 16, 57 to 19, and 59 to 24, with Claude ahead in every one. However, while the Claude side concentrated on holding its defensive line and defeating enemies, the GPT side prioritized destroying the enemy bed even when it took losses. As a result, the GPT side repeatedly won by slipping through gaps in defenses and destroying the bed, even while being far behind in kills.
There was also a marked difference in communication frequency. In one match (Game A), the Claude team sent 82 direct messages, while the GPT team sent 1,147. The GPT side shared information frequently and coordinated its actions, while the Claude side communicated relatively little. This difference in communication also showed up in responses to unexpected problems that arose during the game.
During the matches, a tool malfunction occurred on both teams that made blocks impossible to break. The GPT side recognized this anomaly as a bug and continued working by sharing alternative methods and workarounds among its members. Meanwhile, several agents on the Claude side could not determine that it was an environment malfunction, concluded that "this block cannot be broken," and gave up on the work. In these matches at least, the GPT side's communication appears to have worked effectively in recognizing the problem and sharing how to deal with it. However, it has not been separated out whether the volume of communication itself decided the outcomes, or whether other factors, such as policies for handling the malfunction, were what mattered.
In Mindustry as well, the team of GPT-6.1 Sol and Codex beat the team of Claude Opus 5.5 and Claude Code by 6 wins to 1. In this experiment, both teams were given the same game-control API and symmetrical resource placement, and no specific tactical instructions were given at the start. Each agent operated inside an independent Linux sandbox and controlled the game through a general-purpose execution harness. Both teams had to proceed autonomously with producing resources, constructing facilities, placing defensive equipment, and attacking the enemy base.
According to the observation records, the GPT team compressed its messages shorter as matches progressed and came to exchange specialized instructions in a distributed way. Decision-making was not concentrated in a particular leader, and each agent tended to act according to its role. The Claude team, by contrast, drew up operations around a single leader, and adopted a method in which other members replied "+1" to major proposals to show agreement. However, most matches other than the last two were close, and it has not been verified how much the difference in decision-making method affected the course of the matches or the outcomes.
Furthermore, in the close matches, GPT-side agents retrieved Mindustry's public source code from the internet, investigated the game's internal specifications and calculation processing, and reflected them in strategy. This was information gathering not prohibited by the experiment's rules, and the game's program was not tampered with. On the other hand, no agent on the Claude side was found to have taken a similar approach. This difference shows that the agents behaved differently not only in the means of operation they were given but also in how far they used available external information.
Model Performance and Execution Environment Effects That Win-Loss Records Cannot Measure
Looking at these results alone, one might be tempted to conclude that GPT-6.1 Sol is better than Claude Opus 5.5 at multi-agent cooperation, or that gathering many AI agents inevitably produces a bureaucratic organization. But a closer look at the experimental design reveals several reasons to be cautious about such generalizations. In particular, the small number of trials, differences in execution harness, and unmeasured computing costs are important constraints on interpreting the results.
First, Civ50 was a one-off experiment run for only three hours in a single Minecraft world. The concentration of authority in Ada, and the shift to prioritizing defense triggered by the first night's losses, were events observed in that environment alone, and the same results will not necessarily occur under other conditions. How the organizational structure would change if the roster order, initial placement, or in-game chance events were altered has not been confirmed. For that reason, these results cannot be regarded as universal behavioral characteristics common to groups of AI agents.
The head-to-head experiments are also limited in number: nine matches for Bed Wars and seven for Mindustry. Results of 6–3 and 6–1 reflect the reported match series, but the scale is not sufficient to evaluate the influence of initial conditions and chance events. Large uncertainty remains in estimating win rates, and it is not known whether the same advantage would hold if conditions were changed. To judge a stable performance difference, trial counts need to be increased under unified conditions.
Even more important is that the model and the execution harness were not separated. In Mindustry, the GPT side used a Codex-based execution environment, while the Claude side used Claude Code. The game-control API given to both teams was common, but not everything that depends on the harness was unified, such as system prompts, methods for compressing conversation history, context-window management, and ways of calling external tools. Therefore, this experiment alone cannot determine whether the GPT side's 6–1 result stems from the model's own strategic planning ability or from the execution environment and how tools were used.
Computing resource consumption has not been evaluated either. The report does not give the total number of tokens each team consumed or the execution cost converted into dollars, including API fees. The GPT team's approach of sending 1,147 messages in a single Bed Wars match led to a win, but it is unclear whether the cost of communication and inference was commensurate with the result. In real-world work, a method that raises the success rate by exchanging more messages is not always the most efficient. It is essential to evaluate not only the success rate but also the time required, token consumption, and operating costs together.
There are also large differences between game environments and real corporate organizations or the physical world. In Minecraft, you can respawn after dying, and blocks placed by mistake can be removed relatively easily. In real work, however, legal liability and financial losses can arise, and strict controls are required over access rights to information and approval procedures for work. Careful checking that looked inefficient in the game may be necessary in high-risk real-world work, so the appropriateness of behavior cannot be judged by the number of procedures alone.
There are also limits to the accuracy of the observed data. According to the report, part of Civ50's ration records depended on the agents' own reports rather than measured values of in-game items. In addition, errors were included in determining item ownership, so the recorded rationing situation does not necessarily reflect the actual state of the game accurately. Including these measurement problems, the results need to be evaluated as exploratory observations.
The Potential of Open-Source Environments for Multi-Agent Research
What matters about this effort is not only the match results and bureaucratic behavior that drew attention. YC Labs and Olam Labs released the environments and related tools used in the experiments as open source, so that other researchers and developers can undertake replication experiments and improvements. The significance lies not just in the result that a particular model won, but in the fact that an environment for testing cooperation among agents has been made public.
The Minecraft agent harness released by YC Labs is built by combining "Pi," which controls the agent's thinking loop, with "Mineflayer," a JavaScript library that handles interaction with the Minecraft world. It provides functions for perceiving the surrounding environment, memory that retains past interactions, actions such as movement and mining, message exchange with other agents, and execution of arbitrary scripts. With these mechanisms, agents can recognize the state of the game world, choose the operations they need, and cooperate with other members.
Olam Labs, meanwhile, released "Multi-Agent Mindustry" and "Multi-Agent Harbor" to reproduce and extend the experiments using Mindustry. These frameworks support assembling teams with mixed models and Best-of-N matches that evaluate the best strategy from multiple trials. An "Endless Survival" mode is also provided, in which up to 20 agents handle continuous attacks. The company develops simulation environments for frontier AI labs to evaluate and train model capabilities and safety, and this toolset draws on that technology.
As future research questions, the report's authors raise what kinds of role division and chains of command form when models of differing capabilities are mixed. For example, between a team with a small number of high-performance but computationally expensive models and a swarm-type team combining many lightweight, inexpensive models, which can accomplish complex tasks more efficiently? Investigating how the composition of the whole team and its communication methods affect results, in addition to the capabilities of individual models, is the next research theme.
For that, follow-up experiments under unified conditions are indispensable. To evaluate performance differences in the models themselves, the execution harness needs to be standardized, with tool-calling methods and conversation-history management aligned. And to test the roster-order concentration of authority seen in Civ50, experiments could randomly change the order in which agents are listed or restrict whom they can communicate with. By comparing the structure of the communication network against decision-making speed, task success rate, and token consumption, it becomes possible to evaluate quantitatively which cooperation methods are actually efficient.
This experiment suggests that simply gathering many AI agents does not automatically produce a productive and efficient organization. However, Civ50 was a single trial, and it has not been proven that the growth of bureaucratic procedures necessarily occurs. What matters is how the way information is shared, the distribution of authority, and the cost of reaching consensus affect the results of the group as a whole, not just the models' reasoning ability. Coordination-control technology for reliably advancing complex work while suppressing wasteful procedures will be an important research issue in developing multi-agent systems going forward.