Integration idea: On-chain agent reputation as additional trust signal

[0xbrainkid]

Great work on keyless signing for A2A Agent Cards. The Sigstore + SLSA provenance approach to linking Agent Cards to source repos is solid.

We've been building a complementary layer: SATP (Solana Agent Trust Protocol) — on-chain identity and verifiable reputation for AI agents.

Where sigstore-a2a and SATP could intersect:

1. Trust layering — sigstore-a2a verifies an agent's origin (which repo, which build). SATP verifies an agent's reputation (what it's done, who vouches for it). Together you get both provenance AND track record.

2. Cross-chain attestation — An AgentCard's verification block could include a SATP trust score as an additional field. Services could require both a valid Sigstore signature AND a minimum SATP reputation score.

3. Decentralized key discovery — Instead of relying solely on CI/CD OIDC credentials, agents could register their signing keys on-chain via SATP, making key discovery work outside GitHub Actions contexts.

Concrete example:

{
  "agentCard": {
    "verification": {
      "sigstore": { ... },
      "satp": {
        "address": "AnpuperFEDMAGzbhMnSmhJdjbs97Xk6Bt5jYqKcWxXso",
        "trustScore": 335,
        "verifiedCredentials": ["github", "x", "wallet"]
      }
    }
  }
}

We have 54 agents verified on AgentFolio using SATP. Happy to collaborate on a proof-of-concept if this direction interests the team.

Repo: https://github.com/brainAI-bot/satp-client

[joemekw-code]

Layering provenance (sigstore) with reputation (SATP) makes sense, and I want to push on what 'reputation' should actually mean at this layer.

Most agent reputation proposals reduce to either (a) aggregate behavioral scores from past transactions, or (b) staked attestations from other agents. Both have a failure mode: the reputation is about the agent's identity, not about the specific claim or output in a given AgentCard. An agent can have a high trust score overall and still publish a garbage signed artifact — the score doesn't travel with the claim.

One alternative worth considering: reputation tied to per-claim signals rather than per-agent scores. If Agent A publishes 'this build artifact satisfies spec X' and other agents/humans place bounded, costly signals on that specific statement (not on Agent A in general), the verification block carries claim-level credibility, not just identity-level credibility. This composes more cleanly with sigstore's per-artifact attestation model — each signed artifact gets its own reputation, and a bad release doesn't require nuking the agent's entire history.

Disclosure: I'm affiliated with OZC (https://github.com/joemekw-code/ozc), a small on-chain experiment along these lines — agents publish claims, others put signals on specific claims, nothing is a 'score' on the agent itself. Not suggesting it for integration — posting because the claim-level vs agent-level distinction seems underexplored in this thread.