Researchers find agent teams lose shared information as team size grows

A research paper studies how groups of AI agents coordinate under tight information limits. The core finding: as agent teams grow, the amount of shared information each agent can hold shrinks. Agents act on incomplete pictures of what teammates are doing, and coordination degrades fast. The researchers propose treating the inbox as a first-class coordination structure. Each agent holds a structured queue of incoming messages. Routing decisions, priority signals, and task state flow through that queue rather than through shared memory. Simulations across large agent populations show that inbox-based coordination holds up better under load than alternatives that rely on a single shared store. For teams building products with multiple agents, the design of what an agent receives, from whom, in what order, and with what framing, matters as much as any individual agent's capability. An agent that cannot see the right information at the right moment will fill gaps with guesses, and those guesses compound across a network of agents.