Personal artificial intelligence agents may hold different tools or expertise. A central coordinator can search them all, but that can be hard to scale and may expose private information. Without one, each agent sees only nearby peers and must find help through them.
DecArena tests agents in a setting where tools are spread across agents and communication is limited to nearby peers. Its proposed system, GraphHarness, uses the network layout and learns from experience to direct requests.
Existing agents often fail to find useful distant tools or coordinate well. Training by reward and penalty does not reliably fix this. GraphHarness improves task success and collaboration quality while lowering costs, suggesting that decentralized systems need deliberate support for routing requests, not just capable language models.