New Study Compares DHT Overlays for Decentralized Agent Discovery on 4096 Nodes

As agentic AI systems move beyond centralized clouds into edge and intermittently connected domains, the need for robust, decentralized discovery mechanisms grows. A new paper from researchers Patrizio Dazzi, Emanuele Carlini, Matteo Mordacchini, and Saul Urso, posted on arXiv (2605.11839), directly tackles this challenge. They compare three structured overlay families—Chord, Pastry, and Kademlia—as candidate indexing substrates for agent directories within a shared control-plane framework. The study uses a carefully designed benchmark set: a stationary 4096-node comparison and a representative 4096-node churn benchmark that simulates real-world node joins and departures. Key metrics include discovery reliability (how often agents find each other), startup behavior (time to join and stabilize), and control-plane overhead (bandwidth and message cost).

The results reveal clear trade-offs: Kademlia offers the fastest startup and lowest overhead under light churn, but its reliability degrades significantly under heavy churn. Chord provides more predictable reliability at the cost of higher startup latency and moderate overhead. Pastry strikes a middle ground but struggles with uneven load distribution. The paper's contribution is not a single winner but a map of operating points—helping engineers choose the right overlay based on their deployment environment (e.g., stable cloud clusters vs. highly dynamic edge networks). For practitioners building agentic systems on the compute continuum, this study provides concrete guidance for designing discovery layers that remain effective as agents move across cloud, edge, and intermittently connected zones.