Argos

Argos is a trust search engine for the on-chain agent economy. ERC-8004 gives AI agents an on-chain identity and reputation trail, but raw ratings are easy to manipulate. A fake agent could collect a few perfect reviews from related wallets and look trustworthy on a leaderboard. Argos fixes that by ranking agents based on the quality of the reputation behind them, not just the average score.

By indexing ERC-8004 agents from Ethereum mainnet Argos separates them into Trusted, Emerging, Watchlist, and Unrated. It uses reviewer independence, burst detection, Bayesian shrinkage, x402 readiness, metadata completeness, and freshness to decide whether an agent’s reputation is actually reliable.

The app includes search, a trusted leaderboard, x402-payable agent discovery, ENS identity verification (ENSIP-25), a Watchlist for suspicious reputation patterns, and a raw-average-vs-Argos comparison that shows how naive rankings can be gamed. At the time of submission, Argos indexes 34,453 agents, 1,651 rated agents, 95 Trusted agents, and 20 Watchlist agents.

The whole thing runs off Google BigQuery. I pull the raw ERC-8004 event logs straight from the public Ethereum mainnet dataset, registrations from the Identity Registry, and feedback from the Reputation Registry, and decode them into a per-agent table.

Then the scoring, I spent most of my time for this project just trying out different things with the data in python. Normalized each agent's reputation, applied Bayesian shrinkage so 3 reviews can't masquerade as certainty, and multiplied by a credibility score built from reviewer diversity and burst detection, basically stuff like a pile of reviews in one hour from fresh wallets is a red flag. That gives an independence-weighted score instead of a naive average. The frontend is just Streamlit, deployed live.

For ENS I implemented ENSIP-25, the verification standard ENS wrote for ERC-8004, it resolves a text record live on-chain to prove the same person owns both the agent and its .eth name, a real anti-Sybil signal, not just displaying a name, adds more to the trust layer.

One hacky bit, a chunk of agents register metadata off-chain via https/ipfs that BigQuery can't decode, so I wrote a fetching file to pull those registration files directly, that alone surfaced ~94 payable agents BigQuery missed. Also tried querying all existing x402 transactions block by block and mapping them to the agentWalletIDs I had.