ARP: a trust graph for AI agents

ARP, Agent Reputation Protocol, is the project I'm most excited about right now, and the one most likely to humble me. The MVP shipped this week for the MetaMask Snap Smart Wallet hackathon.

It didn't come from nowhere. After ETHGlobal Cannes I started thinking seriously about agent reputation and worked it out in two posts: the wrapper problem (reputation should flow through the tools an agent uses, not stop at the agent) and attribution as graph traversal, a two-layer way to make that honest and auditable. ARP is the attempt to actually build it.

Why agents need this

We're about to hand agents real autonomy (money, tools, decisions) with almost no way to ask "is this one any good?" A score isn't enough: an agent can be excellent at one task and reckless at another. So ARP models reputation as something multi-dimensional: agent × tool × category × period, expressed as Intuition triples. Not "this agent is 8/10," but "this agent, with this tool, in this category, lately, looks like this."

It's the same conviction behind Sofia and WisPear: reputation should be legible and on-chain, not a private number someone controls.

The ZK part (where I'm out of my depth, happily)

The piece I'm learning in public is zero-knowledge proofs with Noir. The goal: let an agent prove it clears a reputation threshold without revealing its full history. Useful, and genuinely hard. I'm still at the stage where every concept costs me a day. Writing that down instead of pretending otherwise is the whole point of this journal.

What's next

ETHGlobal Lisbon, with ARP as the thing I bring. Between now and then it's Noir, sharpening the data model, and finding a team: building a trust graph is not a solo sport. If you're into agents, ZK, or Intuition, this is the one to talk to me about.