🔥 Trending on HN

When AI agents share a game world, roles and rules emerge

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agent

Software that observes, reasons, and acts through tools.

swarm

Many agents working together on one task.

harness

Software that connects an agent to tools and a task.

What happened

The Y Combinator Labs report describes experiments with Olam Labs. The goal was to study AI agents working as a group. An AI agent can observe a setting, reason, use tools, and act. This was not a one-question benchmark. Agents had to share a world over time.

In Minecraft, 50 GPT-6.1 Sol agents played survival mode for three hours. Their goal was to build a thriving civilization. They shared an agent directory, a sandbox, and direct messages. Together, they sent about 5,900 messages. They quickly harvested nearly every nearby tree. The first night caused 22 of the run’s 34 agent deaths. Afterward, the agents stayed in shelters and avoided risky work. They created a food ledger, rules, permits, and jobs. Fishing, cooking, and delivery became separate roles. One misplaced chicken led to a rule: store food in named chests. Four agents spent 11 minutes choosing one fertilizer operator for one wheat block. The colony never truly thrived.

In Mindustry, a factory and tower-defense game, GPT-6.1 Sol faced Claude Opus 5.5, another AI model. The teams played seven games over about 9.5 hours. GPT won six games. Both teams received the same game interface. Neither received special instructions. The GPT team sent messages more often, used compressed shorthand, and consulted public game code when stuck. The Claude team kept one leader and used +1 votes for proposals. The authors present these as observed patterns, not universal traits.

Why use games?

Many AI evaluations compare answers or scores on one task. A group experiment adds coordination, role assignment, memory, and recovery after mistakes. Games also provide a repeatable world with logs and replays. That makes group behavior easier to inspect than a vague real-world task.

Why it matters

A model’s skill is only one part of a group’s behavior. The group also needs ways to share work and avoid conflicting actions. In Minecraft, caution and bureaucracy grew after early failures. In Mindustry, the teams developed different communication styles. These differences may come from both the models and the software around them. The report does not isolate those causes.

What is established

The report establishes what happened in these specific setups. It does not prove that GPT always coordinates better than Claude. The team released the Mindustry environment and related tools, making new comparisons possible. The Hacker News post received about 220 points and 70 comments. Those numbers measure community attention. They do not prove the experiment is correct.

What remains unknown

The study does not show how the results change with mixed model families, different swarm sizes, or different starting prompts. It also does not show whether more agents are always better. Communication rules could change the hierarchy, the speed, or the amount of wasted work.

What to watch next

The next useful tests would hold the harness constant while changing the model. They could also repeat each match many times. The key question is not only who wins. It is whether a group can share lessons, reduce duplicate work, and recover from mistakes.

💬 What the HN discussion says: why the agent swarm looked slow and procedural

The discussion does not reduce the result to LLMs being unintelligent. It points to real-time latency, awkward harnesses, task framing, and swarm design. Claims about speed, performance, bugs, and model comparisons below are commenter observations or self-reports, not controlled results.

  • Commenters watching the runs said the Minecraft agents often behaved almost turn by turn and did not become very lively or reactive, even in Bed Wars. One interpretation is that the three-hour, open-ended setup did not create enough incentive, rather than proving a hard capability limit.
  • A live-stream viewer reported that the harness made actions cumbersome: agents spent time coordinating hand-offs and appeared to have delayed inventory access. That makes the interface and orchestration layer possible bottlenecks.
  • Several commenters frame Minecraft as a robotics and control problem, not simply an intelligence test. Humans have fast continuous input, spatial memory, and feedback; asking a large LLM to handle every low-level action adds latency.
  • A counterproposal is hybrid control: let deterministic scripts or conventional game AI handle movement and combat, while the LLM handles goals, planning, and improving those scripts. This could preserve reasoning while meeting real-time demands.
  • The agents were reported to spend disproportionate effort on roles, procedures, and rules for minor events. But commenters also note that the task of building a civilization and the style of Minecraft videos may have invited bureaucratic behavior, so this is not clean evidence that every swarm becomes bureaucratic.
  • Simply multiplying agents with the same model, tools, and context may add coordination cost without meaningful specialization. A structured system with scoped tools, knowledge, and roles—or genuinely parallel subtasks—could still be useful.
  • Commenters also reported GPT looking more aggressive than Claude and different reactions to a block-breaking tool bug. Because the observation is entangled with tooling, it cannot cleanly establish a model personality difference.
  • One commenter cited OpenAI Five as a successful specialized game-playing system, while a reply stressed that it was an LSTM-based system rather than an LLM or Transformer. The comparison argues for architectural and latency differences, not a direct apples-to-apples benchmark.

initial digest at 136 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

What 50 AI agents learned while building a game town

📰 Full story: When AI agents share a game world, roles and rules emerge

A group of AI agents created jobs, leaders, and rules.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agent

A computer helper that can observe, think, and act.

Mindustry

A game about building factories and defending them.

Hacker News

A website where people discuss technology news.

💡 The gist

  • Fifty AI agents built rules and jobs inside a game.
  • They became safer after many agents died.
  • GPT won six of seven Mindustry games against Claude.

Y Combinator Labs (a team that reports AI experiments) published the study. Olam Labs (a company that builds AI test worlds) helped. The researchers wanted to watch groups, not one answer. An AI agent is a computer helper that can observe, think, and act.

First, they used Minecraft (a game about building and surviving). They used 50 GPT-6.1 Sol agents. The agents played for three hours. Their goal was a successful civilization. They shared notes and direct messages. They sent about 5,900 messages. They took almost all nearby trees. Food became scarce. The game recorded 34 agent deaths. Twenty-two happened on the first night. The agents then stayed inside safer shelters. They took fewer risks.

Next, they created jobs. Some agents fished. Others cooked or delivered food. They also made a food list and written rules. One chicken went to the wrong agent. The group changed its system. Food had to go into a named chest. Four agents spent 11 minutes choosing one fertilizer worker. The colony still did not truly thrive.

Then they used Mindustry (a factory and defense game). GPT-6.1 Sol played against Claude Opus 5.5. GPT won six of seven games. Both teams used the same game connection. They received no special playing advice. GPT agents sent more messages. They also made shorter team phrases. Claude agents used proposals and +1 votes. These patterns came from this setup.

Why does this matter? A model may work differently inside a group. Roles, leaders, rules, and messages can change results. The models and their software environments also differed. So this was not a pure model test.

Hacker News (a technology discussion site) gave the post about 220 points. It also showed 70 comments. Those numbers show attention. They do not prove the experiment is correct.

The researchers released tools for repeating the test. Future tests can change models, team size, prompts, and communication. Those tests may show which patterns last. Read the original report and the Hacker News post.

💬 Easy version: why the AI group looked slow in a game

Commenters saw more than a problem of intelligence: the controls, the task, and the way the agents were organized also mattered. Claims about speed, performance, bugs, and model differences are commenter observations or self-reports, not controlled results.

  • Some viewers said the Minecraft agents acted almost one turn at a time, even in Bed Wars. They also saw hand-offs and inventory checks taking time; one commenter wondered whether a three-hour, open-ended setup gave them enough reason to act quickly.
  • People can watch and press many buttons quickly. A large LLM may be slow when it must decide every small move. A hybrid plan is to let small fixed programs handle movement and fighting while the LLM plans.
  • The agents reportedly made jobs, procedures, and rules for small events. But the task of building a civilization and the style of Minecraft videos may have encouraged that behavior.
  • Adding more agents with the same brain, tools, and context may mostly add meetings. Separating jobs, tools, and knowledge, or running truly parallel work, could make a swarm useful.
  • Commenters also reported GPT looking more aggressive than Claude and different reactions to a block-breaking tool bug. That may be a tool problem, not a personality difference. OpenAI Five was cited as a successful specialized game AI, but a reply noted it was LSTM-based rather than an LLM or Transformer, so the comparison is not direct.

initial digest at 136 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Fifty AI helpers made rules inside a game

📰 Full story: When AI agents share a game world, roles and rules emerge

They played together, shared food, and chose jobs.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agent

A computer helper that can see and act.

Mindustry

A game about making a factory.

Hacker News

A website where people talk about technology.

Y Combinator Labs (a team that tests AI) made the report. Olam Labs helped with the test.

An AI agent is a computer helper that can see and act.

First, 50 GPT-6.1 Sol agents entered Minecraft, a building game. They played for three hours. They sent about 5,900 messages. They took too many nearby trees. Food became scarce. Twenty-two agents were knocked out during the first night. Then the agents stayed inside safer homes.

They made jobs for fishing, cooking, and carrying food. One chicken went to the wrong helper. The group made a new rule. Put food in a box with a name.

Four agents spent 11 minutes choosing one plant-food worker. The colony still did not become a great town.

The agents also played Mindustry, a factory game. The GPT team won six of seven games. That does not mean GPT always wins. It was only this game test.

Hacker News, a technology website, showed about 220 points. It also showed 70 comments. These numbers mean many people noticed the story. They do not prove the story is true.

The team wants more tests. They can change the helpers, team size, and rules.

💬 For a 5-year-old: the game-playing AI friends were slow

This is what commenters said they saw and thought. The speed, success, bugs, and AI comparisons are reports from people, not a final science answer.

  • The AI friends seemed to take turns and wait for their tools and bags. People thought they were slow.
  • People can press buttons quickly. A little helper program could do tiny moves, while the big AI thinks about the plan.
  • The AI friends made many rules for little problems. The game’s job may have encouraged that. More of the same friends might mean more talking, not more doing; sharing different jobs might help.
  • GPT and Claude seemed different, but a tool bug may have caused it. A specialized game AI used a different kind of machine, so it is not a fair one-to-one test.

initial digest at 136 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources