CoAct: A Minimal Coordination Substrate for Decentralized LLM Agents

Large language models are increasingly deployed as tool-using agents, motivating multi-agent systems that can explore in parallel, share discoveries, and cross-check decisions. However, decentralized collaboration introduces two core challenges: agents must make local discoveries available to the team, and they must coordinate before executing stateful actions that modify shared or externally visible environments. We introduce CoAct, a minimal coordination substrate for decentralized LLM agents based on two primitives: compact shared-memory communication and synchronized stateful commitment. CoAct preserves the standard ReAct-style agent loop while allowing agents to independently use stateless tools, write decision-relevant messages to a shared memory, and jointly select stateful actions before execution. Across four agentic task domains spanning math reasoning, web research, financial analysis, and stateful user assistance, CoAct improves over single-agent execution, non-communicating parallelization, and multi-agent system with unstructured communication, while using communication more token-efficiently than free-form discussion. We further show that scaling beyond small teams leads to non-monotonic advantage: increasing the number of agents or the capability of the base model does not always improve performance. Through empirical scaling studies, probabilistic analysis, and case studies, we identify when communication helps or hurts decentralized agent teams, suggesting that progress in multi-agent LLM systems depends not only on scaling teams but also improving effective communication.