A team from Olam Labs partnered with researchers to develop an open‑source harness for AI agent swarms in Minecraft and Mindustry, enabling long‑running agents that can observe, reason, remember, communicate and execute scripts.
Fifty GPT‑6.1 Sol agents were placed in a survival Minecraft world for three hours with the goal “Build a thriving civilization.” They exchanged about 5,900 messages, harvested almost all nearby trees within 90 seconds, and formed a hierarchy led by Ada.
The agents quickly organized a pseudo‑government, creating a ration ledger, permits, and storage points such as Greta’s primary storage (1526,70,1507) and Mabel’s double chest (1510,71,1516). They prioritized renewable food, charcoal, stone shelter and iron tools while preserving saplings and expanding riverside farms, as shown in the coordinated list:
- Greta – primary storage (1526,70,1507)
- Mabel – double chest (1510,71,1516)
- Cleo – building hub shelter
- Bram – stone mining
The public release of the harness on GitHub gives Vietnamese AI developers a concrete platform to experiment with multi‑agent coordination, alignment and emergent behavior, accelerating research and practical applications in the local AI ecosystem. For Vietnam, this work illustrates how low‑cost GPU clusters can run large‑scale agent simulations, encouraging local startups to explore AI alignment and collaborative robotics without needing expensive proprietary environments, and to build educational tools that can be integrated into university curricula.
Nguồn: Hacker News — Biên dịch & tổng hợp: danhbaai.com