Multi-agent AI learns to ignore stale messages — first metric for measuring when communication helps versus hurts

Researchers formalized a problem that happens in real multi-agent systems: messages arrive late, so agents make decisions on outdated information. They built a measurement (CGDC) that quantifies when a delayed message actually helps coordination versus when it just adds noise, then used it to train agents to request messages only when useful. In practice, this means multi-robot teams and distributed AI systems can now decide whether to wait for late information or act on what they know, instead of treating all communication as equally valuable.

For years, multi-agent reinforcement learning systems treated communication like it's always good — more information, better coordination. But in real deployments, messages get delayed by network latency, processing time, or physical constraints. The practical result is agents coordinating on stale data, which can actually make performance worse than no communication at all. This paper shows how to measure the trade-off mathematically and use it to make smarter decisions about which messages are worth waiting for. That matters because it's the difference between a robot team that waits for perfect information and fails versus one that commits to incomplete information and succeeds.