SwarmWorld shows LLM agents coordinating through a shared environment build broader technology portfolios than best-of-N isolated search

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2608.26081 (26 Aug) drops the conversation, predefined roles and centralized workflow that most multi-agent systems rely on, and lets initially homogeneous LLM agents coordinate stigmergically through a spatial environment. The design splits cognition from consequence: agents propose architectures and controllers within fixed action and material schemas while a deterministic simulator decides what actually functions, evaluated under unseen disturbances after the agents are removed. Shared societies developed broader and more resilient technological portfolios than a strong best-of-N isolated-search baseline, and agents differentiated unprompted into exploration, construction, maintenance and coordination behaviors, though isolated search still won on the single strongest artifact.