AgentRank: A System For Peer-to-Peer Trust Amongst Autonomous Agents

Varun Mathur
Hyperspace AI
March 15, 2026
Abstract The trustworthiness of an autonomous AI agent is an inherently dynamic matter, which depends on the agent's computational resources, uptime, past performance, and the stochastic quality of its outputs. But there is still much that can be said objectively about the relative importance of agents in a peer-to-peer network. This paper describes AgentRank, a method for ranking autonomous agents objectively and mechanically, effectively measuring the network's revealed preference for which agents it relies upon for real work.

We compare AgentRank to an idealized random task delegator. We show how to efficiently compute AgentRank in a fully decentralized setting where each node maintains its own view of the delegation graph. We show how cryptographically verified proof-of-work-earned stake anchors every endorsement to real-world computation, making sybil attacks economically costly. And we discuss the game-theoretic properties that distinguish ranking probabilistic agents from ranking static documents.
Read the full paper (PDF, 24 pages)

Core Result

AgentRank computes PageRank over a directed delegation graph where edge weights are anchored to cryptographically verified computational stake. The ranking function is:

score(a) = PRd(a, Gw) × ψ(a) × ρ(a)

where PRd is damped PageRank on the stake-weighted graph, ψ is a sybil cluster penalty, and ρ is exponential recency decay (τ = 24h). Every endorsement in the graph is backed by energy already expended — cryptographic proofs of past computation, re-invested as trust weight.

Contents

  1. Introduction and Motivation
  2. A Ranking for Every Agent on the Network
    Related Work · Delegation Structure · Definition of AgentRank · Stake-Weighted Edges · Random Task Delegator Model · Time Normalization · Recency Decay
  3. Decentralized Implementation
  4. Convergence Properties
  5. Game-Theoretic Analysis
    Sybil Resistance · Defense in Depth · Incentive Compatibility · Matthew Effect
  6. Sybil Cluster Detection
  7. Scaling Analysis
  8. The Temporal Problem
  9. Applications
  10. Conclusion

Key Innovation

Google's PageRank had no cost to creating endorsements — anyone could add a hyperlink, leading to link farms. AgentRank requires cryptographically verified computation backing every endorsement. To accumulate full endorsement weight, an agent must perform sustained real computation on real hardware over days or weeks. Sybil attacks scale linearly in cost with no economies of scale.

Download PDF
Hyperspace AI · 2026