We Tested AI Agent Swarms in Minecraft and Mindustry
agents claude
| Source: HN | Original article
Researchers deployed swarms of AI agents in the games Minecraft and Mindustry, observing distinct multi‑agent behavior patterns across the simulations.
Researchers from YC Labs and Olam Labs have released the first public results of large‑scale AI agent swarms operating in two sandbox games, Minecraft and Mindustry. In a three‑hour “colony” experiment, 50 GPT‑6.1 Sol agents were dropped into a vanilla Minecraft world. Over the run the agents exchanged roughly 5,900 messages, organized themselves into a hierarchy centred on the first‑listed name, and became noticeably more risk‑averse after 22 of the 34 agents died during the first night.
The same GPT‑6.1 Sol swarm was pitted against a team of Opus 5.5 agents in Mindustry, a tower‑defence‑style production game. The Sol team secured a 6‑1 victory, demonstrating that the agents could not only gather resources and build structures but also coordinate defensive strategies against wave attacks.
Why it matters is twofold. First, the experiments provide concrete data on how dozens of language‑model‑driven agents negotiate roles, share information and adapt to setbacks in open‑ended environments—behaviours that are central to the emerging field of multi‑agent AI alignment. Second, the distinct patterns observed in Minecraft’s free‑form crafting versus Mindustry’s production‑line constraints highlight how task structure shapes emergent coordination, informing future designs of AI systems that must collaborate on real‑world problems.
The teams plan to extend the work with a new 25‑agent colony of GPT‑6.1 Terra bots, streaming the runs on Twitch. Observers will be watching for whether similar hierarchies and risk‑aversion dynamics arise, and how the agents handle more complex, mixed‑objective scenarios. Continued monitoring will be crucial for gauging both the capabilities and safety implications of large‑scale AI swarms.
Sources
Back to AIPULSEN