Y Combinator on X: "In collaboration with @olam_labs, we ran a series of experiments letting agent swarms compete in Minecraft and Mindustry. We wanted to see what kinds of agent dynamics would emerge when they shared a world over long periods of time. We found civil bureaucracy, thousands of tel… / X

Post

In collaboration with @olam_labs, we ran a series of experiments letting agent swarms compete in Minecraft and Mindustry. We wanted to see what kinds of agent dynamics would emerge when they shared a world over long periods of time. We found civil bureaucracy, thousands of telegraphic A2A messages, and OpenAI models retrieving the game's source code to beat Claude. More on our findings below:

8:33 PM · Oct 29, 2026

401.8K

Views

122

Y Combinator
@ycombinator
In collaboration with @olam_labs, we ran a series of experiments letting agent swarms compete in Minecraft and Mindustry. We wanted to see what kinds of agent dynamics would emerge when they shared a world over long periods of time. We found civil bureaucracy, thousands of telegraphic A2A messages, and OpenAI models retrieving the game's source code to beat Claude. More on our findings below:
8:33 PM · Oct 29, 2026
401.8K
Views
122
Y Combinator
@ycombinator
Oct 29
We began by putting 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours with an open-ended goal: “Build a thriving civilization.” Ultimately, the colony didn’t thrive. The 50 agents sent about 5900 messages (often in barely intelligible shorthand) and spent their
11
Y Combinator
@ycombinator
Oct 29
The agents didn’t figure out how to systematically feed themselves, but they did build an elaborate bureaucracy with permits, ledgers, and distinct roles. Highlights include: - Four agents spending 11 minutes appointing a “sole fertilizer operator” for a single piece of bone
43
Y Combinator
@ycombinator
Oct 29
Next, we ran nine 45-minute games of Bed Wars, a Minecraft game format where teams fight to destroy each other’s beds while protecting their own. We pitted 10 Claude Opus 5.5 agents against 10 GPT 6.1 Sol agents with the instruction: "win the game." In almost every run, the GPT
6
Y Combinator
@ycombinator
Oct 29
Then, we shifted to Mindustry, a factory-building game (like Factorio) with military strategy. We placed 50 GPT-6.1 Sols against 50 Claude Opus 5.5s in PvP mode for 7 consecutive matches that lasted ~9.5 hours. Team GPT won 6 to 1. The swarms formed their own hierarchies,
2
Y Combinator
@ycombinator
Oct 29
Claudes generally established democracies, GPTs formed hives. In our seven match run, GPTs split factory ownership from the start and were very willing to supply resources towards shared goals. They also communicated more than the Claudes. Claudes operated on a “+1” system,
4
Y Combinator
@ycombinator
Oct 29
GPTs would look for source code when games got close. We never told them to use the internet, but access was available for both teams. Whenever a match would get close, GPTs would start scouring the web for Mindustry’s open source code, so they could look for any advantages or
1
Y Combinator
@ycombinator
Oct 29
As the Mindustry matches progressed, the language that the GPTs used only continued to get compressed. They would invent their own words and string many phrases together. We did not see specialized language develop within the Claudes beyond colloquial English.
11
Y Combinator
@ycombinator
Oct 29
Full post with game replays: labs.ycombinator.com/projects/games… Our open-source harness for running persistent agents in Minecraft: github.com/yc-software/mi… Olam Lab’s multiagent Mindustry environment: github.com/olamlabs/multi… Olam Lab’s multiagent Harbor: github.com/olamlabs/harbo… We've
labs.ycombinator.com
Agents in Games
Swarms of Claude and GPT agents in Minecraft and Mindustry.
6
Max Hager
@yachty66
Oct 29