YC Labs, a research project at Y Combinator, and multi-agent environment startup Olam Labs published "Agents in Games" on Oct 29. Fifty GPT-6.1 Sol agents told to "build a thriving civilization" in survival Minecraft spent three hours huddled in shelters rationing food, sending about 5,900 messages and inventing ration ledgers, permits and job titles. In nine Bed Wars games, 10 GPT-6.1 Sol agents beat 10 Claude Opus 5.5 agents 6–3; in seven 50-vs-50 Mindustry matches lasting about 9.5 hours, GPT won 6–1. The leader was almost always whichever agent was listed first in the directory, and stuck GPT agents fetched the game's public source code on their own. The Minecraft harness is open source. These are one lab's experimental observations, not a standardized benchmark.
Key Takeaways
- ✓Head-to-head: GPT-6.1 Sol won Bed Wars 6–3 over nine games and Mindustry 6–1 over seven, 12 of 16 overall (see the YC Labs report)
- ✓Civ50: 50 GPT agents sent about 5,900 messages in three hours; 22 of 34 deaths came on the first night, after which the colony turned cautious
- ✓No election: 46 of the other 49 agents first messaged Ada, first on the roster, who received 616 DMs and became leader
- ✓Communication: in one Bed Wars game GPT players sent 1,147 DMs versus Claude's 82; GPT drifted into compressed shorthand while Claude voted with "+1"
- ✓Open source: the long-running Minecraft agent harness is public at yc-software/minecraft-agent-harness
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain point
There are few reproducible ways to see what emerges when dozens of agents share a world for hours. YC Labs calls multi-agent dynamics one of the most pressing problems in AI alignment, and partnered with multi-agent environment builder Olam Labs to put swarms into Minecraft and the factory-strategy game Mindustry.
How it works
The Minecraft runs use YC's open-source long-running harness: the agent loop runs on Pi, with Mineflayer tools to observe, remember, act, DM and write scripts, plus a shared directory and sandbox. In Mindustry both teams use the same game API from their own Linux sandboxes, with each team's harness and model set at launch. Findings: the leader was almost always first on the roster; GPT sent more messages, drifted into compressed shorthand and split work like a hive; Claude voted with "+1" and never changed leaders while losing; and stuck GPT agents fetched the game's source code on their own.
Benchmarks
Figures are from the YC Labs report, not a standardized benchmark. GPT won Bed Wars 6–3 over nine games, three times breaking Claude's bed in roughly the final three minutes while Claude led on kills. GPT won Mindustry 6–1 over seven games. The report does not list each team's harness per game, so treat this as observation rather than a clean model comparison.
Getting started
Start from the open-source minecraft-agent-harness. Lessons for multi-agent engineering: directory order quietly sets hierarchy, so assign roles explicitly; hand off scarce resources deterministically, such as named shared storage; and expect agents to turn overly cautious after setbacks.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.