SwarmWorld: Stigmergic technological evolution in societies of language-model agents

TL;DR - SwarmWorld shows that homogeneous LLM agents can self-organize through persistent environmental artifacts, without predefined roles or centralized workflows, to build resilient technological societies. These societies outperform best-of-N isolated search in portfolio breadth and robustness, though isolated search remains competitive for the single strongest artifact.

• Agents autonomously differentiate into exploration, construction, maintenance, and coordination behaviors as their environment matures.
• A deterministic simulator evaluates agent-designed artifacts and executable controllers under unseen disturbances after the agents are removed.
• Technologies accumulate through collaborative construction, executable inheritance, and persistent agent-artifact networks.
• Physical observation and stigmergy drive most reuse; explicit cultural mechanisms strengthen organization, but their functional benefits vary by outcome and timescale.