We ran AI agent swarms in Minecraft and Mindustry
TL;DR Highlight
We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.
Who Should Read
A backend developer designing a multi-agent system who wonders whether adding more agents actually improves collaboration
Core Mechanics
- 50 agents built a government
- 50 GPT-6.1 Sol agents ran for 3 hours in a survival world with the goal "Build a thriving civilization."
- They created a ration ledger, written rules, permits, and appointed operators.
- They sent about 5,900 messages.
- Directory order became the hierarchy
- Ada, first in the agent directory, became the de facto leader.
- Other agents messaged her for approvals.
- She wrote a colony plan in the shared sandbox for everyone to read.
- First-night deaths changed behavior
- 22 of the run's 34 deaths happened on the first night.
- Death only cost gear, yet agents huddled in shelters on later nights.
- They attempted fewer risky tasks.
- Goal shifted from thriving to food rationing
- 50 agents meant scarce food, tracked in food-queue.md.
- Ines handled requests and chest in/out tracking.
- After the third night Ada told agents to stop optional excavation, building and scouting.
- The harness is open source
- The core agent loop runs on Pi, with Mineflayer-based tools for world interaction.
- It was inspired by repos like Mindcraft.
- The repo is github.com/yc-software/minecraft-agent-harness.
Community Reactions
- Prompt order decided the leader, per one commenter
- A commenter noted 46 of 49 agents messaged Ada first, and she was first in the list.
- They suggested shuffling the order per agent to expose harness bias.
- LLM latency is the bottleneck, per commenters
- One commenter said 500ms of latency per mouse movement makes fighting a spider scary.
- They argued Minecraft is a robotics-style continuous feedback problem, not an intelligence problem.
- A three-tier setup for playing games with Claude
- One commenter played Graveyard Keeper 2 with Claude Max and said it was a blast.
- Fast decisions ran in an in-game mod, macro decisions on a policy server, and Claude connected via MCP.
- They said it saves tokens on mundane tasks.
- Pushback on 50-agent setups
- Some commenters argued 50-agent code factories waste tokens versus one agent.
- Others asked whether swarms beat more structured approaches.
How to Apply
- Check agent-order bias first
- In a multi-agent prototype, shuffle the agent list order on each run.
- Check in the logs whether the leader role keeps going to the same agent.
- Make shared state a single file
- Have agents write status and update times into a shared markdown file such as food-queue.md.
- Make them separate reported values from confirmed values.
- Add duplicate-handling rules
- For tasks that split scarce resources, put a reserve-then-confirm step in the prompt.
- Forbid marking items as done until the recipient confirms.
- Compare with fewer agents
- Run the same task with one agent and with several.
- Compare token cost and results, and count coordination messages in the logs.
Try It
# Paste into Claude or ChatGPT — simplified version of the paper's ration-ledger approach
You are an agent playing the role of {{role}}. Shared resources are {{resource list}}.
Rules:
1. Before using a resource, write a reservation request in the ledger.
2. In the ledger, separate 'reported values' from 'values the recipient confirmed personally'.
3. Do not mark anything 'done' until the recipient confirms.
Update the ledger for the situation below and decide the next action: {{current situation}}
Next step: give the same situation to 3 agents and compare whether duplicate pickups occur.Terminology
Related Papers
AI isn't replacing SREs, it's deskilling them
Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.
Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor
Building an AI-native engineering org
Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%
The Agent SDK
Cursor released Agent SDK, letting you define agents in code and deploy them with one command.
Here be monsters: What do you still need to know when agents do the work?
Owning agent-written code depends on fundamentals, breadth, and diving deep just in time.
AI promised less work. I've never worked more
Running agents in parallel brought my GitHub activity back to full-time engineer levels, and I'm exhausted.
Show HN: Corgi, a helpful {agent, terminal} multiplexer
A tmux-like terminal tool that shows the status of all your AI agents at a glance
Ask HN: If the models are so good, why doesn't OpenAI just start 100 companies?
HN answers: code is easy, knowing what to build isn't, so they sell tokens instead