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

gist

In SwarmWorld, initially identical language-model agents explore a spatial world, process resources, build persistent artifacts, and write executable controllers. A deterministic simulator, not the agents' claims, decides what works, including after the agents are removed. Shared societies form broader, more resilient technology portfolios than a strong isolated-search baseline, though isolated agents can still produce the single strongest artifact. Most reuse begins by observing the world rather than talking; physical stigmergy alone supports capable societies.

ideas

• Split cognition from consequence. Agents propose architectures and controllers inside fixed schemas; the simulated world determines function, including under unseen disturbances after the agents are gone.
• Shared worlds win on portfolios, not always on the best object. Societies beat a strong best-of-N isolated-search baseline on breadth and resilience, while isolated search remains competitive for the strongest artifact.
• Physical stigmergy is enough. Most reuse begins by observing persistent artifacts rather than communicating; explicit culture amplifies collaboration, but functional benefits depend on outcome and timescale.
• Unassigned agents differentiate. Exploration, construction, maintenance, and coordination emerge without assigned roles or recipes, and those behaviors shift as the world matures.
• Interaction builds an ecology. Technologies accumulate through collaborative construction, executable inheritance, and agent-artifact networks rather than universally superior individual inventions.

quotes