Usable Agent Discovery for Decentralized AI Systems

A new research paper from Patrizio Dazzi, Emanuele Carlini, Matteo Mordacchini, and Saul Urso (arXiv:2604.23080) dives into the challenge of usable agent discovery in decentralized AI systems. As large-scale agentic systems run on distributed infrastructures, software agents must share physical hosts and discover each other via peer-to-peer mechanisms. The key twist is two-level churn: node-level churn from failures and host departures, and agent-level churn from demand-driven activation, deactivation, and state changes (e.g., warm vs cold states). The authors compare structured overlays (Kademlia) against gossip-based overlays (Cyclon+Vicinity) under stable, node-churn-only, agent-cooling-only, and combined regimes.

The findings are nuanced: structured overlays are more robust and efficient in stable and node-churn regimes, while gossip-based overlays remain competitive and can be faster when service readiness dominates. This research reshapes classic trade-offs between structured and unstructured overlays, offering practical guidance for engineers building decentralized AI systems. For tech professionals, this means that the choice of discovery mechanism depends heavily on the expected churn patterns—structured overlays for reliability under failures, gossip for speed when agents frequently switch states. The work is a must-read for anyone architecting large-scale, distributed agentic systems where resilience and latency are critical.