A new empirical study from University of Pisa and CNR ISTI compares Kademlia based routing against gossip protocols for AI agent discovery, finding that Kademlia excels in stable infrastructure environments while gossip protocols outperform it when agents exhibit rapid…
The Linux Foundation is making a decision right now that will determine how every AI agent on the internet finds every other AI agent. The bet is Kademlia: a structured overlay protocol that routes queries along fixed logical paths, like a phone directory for machines. The alternative is gossip — agents telling each other who is online, the way humans spread word of mouth. A new empirical study from University of Pisa and CNR ISTI researchers finally puts numbers on which approach holds under real-world agent behavior. The finding: it depends entirely on how agents are used. And that is exactly the problem.
AGNTCY — the Linux Foundation project backed by Cisco, Google Cloud, Oracle, Red Hat, Dell, and 75 other companies — built its Agent Directory Service on Kademlia. AGNTCY is the most serious attempt yet to create a universal registry for AI agents, a kind of DNS for the machine economy. If you're building a multi-agent system today and want other agents to find yours, AGNTCY is where you register. Its architecture is Kademlia all the way down: a distributed hash table that routes queries along fixed logical paths determined by node IDs.
The Pisa study tested both approaches under four distinct regimes: stable loads, node-level churn, agent-level cooling cycles, and combined churn. The results: Kademlia wins when the network is stable or when only nodes are failing. But when agents are constantly activating and deactivating — cooling down after completing tasks and spinning back up for new ones — gossip protocols pull ahead. The crossover point is not theoretical. The paper gives builders the parameters.
The reason is intuitive once you see it. Kademlia maintains routing state that goes stale when agents disappear and reappear. Discovery might succeed technically but arrive too late to meet a task deadline. Gossip protocols propagate readiness signals as part of their normal chatter — every node that knows an agent is online also propagates that fact. The paper introduces a metric called U_Δ, or useful availability: not just whether an agent can be found, but whether it can be found quickly enough to matter.
The Pisa team — Patrizio Dazzi and Saul Urso from the University of Pisa, Emanuele Carlini and Matteo Mordacchini from CNR ISTI — ran their simulation with 4,096 logical agents on 2,048 physical hosts, modeling the churn patterns that emerge when agents run as serverless functions in response to workload. That is how most cloud-native AI applications are built. It is also not the regime Kademlia is optimized for.
The paper does not name AGNTCY. But its experimental design is a direct test of AGNTCY's assumptions. The regime where Kademlia wins — stable infrastructure, moderate node churn — describes a carefully managed data center. The regime where gossip wins describes what happens when agents are deployed the way the industry is actually deploying them.
The deeper issue is governance. Structured overlays concentrate routing authority. A Kademlia network routes queries along fixed logical paths determined by node IDs. If you control the ID assignment, you control the routing topology. Gossip protocols distribute that authority — every node propagates what it knows, nobody owns the global state. The Pisa paper is not about governance. But its findings illuminate a choice that standards bodies are making right now, largely without public debate: who controls how agents find each other, and whether that control should be concentrated in a topology or dissolved into protocol chatter.
AGNTCY's documentation makes the DNS analogy explicit. That is a reasonable bet if the dominant workload is stable. The Pisa paper says it is not obviously correct for the workloads that are actually taking over. AGNTCY has been running since March 2025, donated to the Linux Foundation in July. It has real code and real adoption pressure. But its operators have not published telemetry on agent churn rates, cooling frequencies, or discovery latency under production load.
Independent researchers see the gap. A team from Lancaster University and the University of Queensland published a complementary analysis arguing that gossip protocols represent a missing layer for emergent, context-rich agent coordination that structured approaches cannot easily replicate. Their conclusion converges with the Pisa findings: temporal staleness, semantic filtering, and trustworthiness in peer-to-peer state exchange are unsolved problems that benchmark studies alone cannot resolve.
A technical analysis published on DEV Community by Dr. Hernani Costa laid out the protocol stack clearly: MCP handles tool integration, A2A handles agent delegation, ANP (AGNTCY's underlying protocol) handles cross-network discovery. Costa's framing — that ANP's value proposition is decentralized discovery without a central directory — is accurate. But it sidesteps the performance question the Pisa paper finally puts numbers on.
The paper is a simulation study. Its findings depend on parameters that real AGNTCY deployments might tune around. Kademlia implementations can be extended with caching, warm pools, and application-layer readiness hints. The authors may have conservatively tuned parameters an experienced systems engineer would adjust. These are not dismissals — they are the exact questions AGNTCY's maintainers should be answering publicly.
What the paper provides that did not exist before is an empirical regime map. For the first time, there are numbers on where the trade-offs flip. Builders making infrastructure decisions now have something to point to beyond intuition and marketing. Whether AGNTCY's architects treat it as relevant data — or treat it as an academic exercise run on parameters they have already solved — is itself the measure of whether the Internet of Agents is actually being built with open eyes.