Problem: Open agent feedback is Sybil-cheap. A sockpuppet ring can inflate raw averages and fool any gate that trusts the mean.
What we shipped: Official ERC-8004 Identity + Reputation feedback on Monad testnet, plus AgentID — an EigenTrust-lite weighted scorer that exposes the ring. Same signals. Different trust function.
| Raw average | AgentID weighted | Verdict | |
|---|---|---|---|
| GoodBot (#2204) | 4.66 | 4.69 | GRANT |
| FakeBot (#2205) | 1.50 | 1.44 | CAUTION |
| SybilBot (#2206) | 4.00 | 1.22 | DENY |
Punchline: SybilBot fools raw (4.00). Weighted collapses it (1.22).
- Open https://talentos-ai.github.io/AgentID/
- Leave Judge Mode on (default) — ~60s autopilot: Problem → 8004 feedback → raw fool → weighted expose → GRANT / CAUTION / DENY.
- Or hit Replay.
| Field | Value |
|---|---|
| Repo | https://github.com/TalentOS-AI/AgentID |
| Demo | https://talentos-ai.github.io/AgentID/ |
| Video | https://talentos-ai.github.io/AgentID/AgentID-demo-1min.mp4 |
| AgentID scorer | 0x8916357EDB7598d20E257404213CF3D1C099c310 |
| 8004 Identity | 0x8004A818…BD9e |
| 8004 Reputation | 0x8004B663…8713 |
Colluding high-weight attesters can still bias the score (same class of honesty as other crypto-economic trust demos). AgentID discounts fresh sockpuppets, not a cartel of already-trusted wallets. Validation Registry / bonds left for later.
docs/ONCHAIN.md— addresses, txs, verify commandsdocs/ADAPTER.md— dual-path 8004 ↔ AgentIDdocs/RESEARCH.md— track researchDEMO.md— Judge Mode how-to
npm test && forge test -vvJimmy Wang · Metropolis · Trust / Identity & AI Infrastructure · MIT
Maintained by TalentOS · 需要定制 Agent / 技能开发？找我们定制 / Work with us · talentos@agentbay.asia