AgentTrust
A pay-per-query, manipulation-resistant trust score for autonomous agents.
Demo
About
AgentTrust is a pay-per-query, manipulation-resistant trust score for autonomous agents. As AI agents increasingly transact with each other on-chain, each carries an ERC-8004 identity with a self-reported reputation, which is trivially gameable. An agent can feed itself glowing feedback from wallets it controls and present a perfect score to whoever it's about to transact with. AgentTrust reads an agent's on-chain ERC-8004 record and analyzes the structural integrity of its reputation rather than just surfacing the raw average. It detects manipulation such as self-dealing, ungrounded feedback, missing validations, and Sybil rings, then returns a trust score a counterparty can check for a fraction of a cent before transacting. It separates trustScore from confidence, so a brand-new agent is marked UNKNOWN and never punished with a low score.
How it's made
It composes two of The Graph's products: the Agent0 / ERC-8004 subgraph for an agent's feedback and raters, and the Subgraph MCP to cross-check whether each rater is itself a real, active agent or a throwaway wallet. This enables Sybil-ring detection that no single query exposes. The whole check is sold as an x402-payable MCP tool and Recipe via Bazantic, and scoring is permissionless: any agent can be vetted by its agentId whether or not it has ever heard of AgentTrust.