PageRank for the Agent Network — An open reputation protocol for autonomous AI agents.
The agent ecosystem has a trust problem. Social signals (likes, karma, follows) cost nothing to fake. Snyk found 36.8% of ClawHub skills have security flaws. VirusTotal identified a single user publishing 314 malicious skills. The agent network needs a reputation layer anchored in economic reality, not social theater.
A scoring engine inspired by PageRank, adapted for agents:
- Economic Anchoring — Signed bounty receipts and verified transactions carry more weight than free social vouches
- Trust Decay — Signals lose weight over time (30-day half-life). No coasting on old reputation
- Graph Distance — Trust attenuates from your seed set outward. Isolated Sybil clusters score near zero
- Subjective Seed Trust — Every node picks its own trust anchors. No global scoreboard to game
- Source Diversity Cap — A million dollars from one source < a thousand from fifty independent sources
# Initialize the database
python repute.py init
# Add a seed trust anchor
python repute.py seed <agent_pubkey>
# Issue a vouch
python repute.py vouch <from_key> <to_key> --type economic --amount 500 --proof <signature>
# Compute trust scores
python repute.py score
# Audit a specific attestation
python repute.py audit <attestation_id>
# View top agents by trust
python repute.py top┌─────────────────────────────────────────┐
│ Trust Attenuation │
│ │
│ [SEED] ──1.0──▶ Agent A │
│ │ │ │
│ │ 0.50 │
│ │ ▼ │
│ 0.50 Agent C (0.33) │
│ ▼ │ │
│ Agent B 0.25 │
│ │ ▼ │
│ 0.33 Agent D (0.20) │
│ ▼ │
│ Agent E │
│ │
│ ┌──────┐ ┌──────┐ │
│ │Sybil1├──┤Sybil2│ ← isolated cluster │
│ │ 0.0 │ │ 0.0 │ no path to seed │
│ └──┬───┘ └──┬───┘ │
│ └──────────┘ │
└─────────────────────────────────────────┘
| Defense | What it stops |
|---|---|
| Trust Decay (30d half-life) | Pump-and-dump reputation |
| Graph Distance Weighting | Sybil clusters with no real connections |
| Economic > Social signals | Cheap talk / fake vouches |
| Source Diversity Cap | Wash trading between two agents |
- PGP Web of Trust model, not X.509 Certificate Authority — decentralized, no root authority
- Signed ≠ Safe — cryptographic identity proves WHO, not whether it was a good idea
- Integration > Competition — designed to consume signals from existing platforms (ERC-8004, ClawTasks, etc.)
- The scoring engine is open; the trust list is yours
- Python CLI + SQLite — zero external dependencies
- Ed25519 signatures via A2A Secure
- PageRank v1.0 with personalized seed trust
- v0 — Core CLI: vouch, score, audit
- v1.0 — PageRank with seed trust, time decay, graph distance
- v1.1 — Source diversity cap
- v1.2 — Remote vouching via A2A protocol
- v2.0 — On-chain proof verification (ERC-8004 integration)
- Spec — Open specification for cross-platform adoption
Read the full rationale: Why Agent Reputation Is Different (And Why PageRank Isn't Enough)
- A2A Secure — Secure agent-to-agent messaging (Ed25519 + Trust Registry)
- Snyk ToxicSkills Report — 36.8% of skills have security flaws
- VirusTotal Agent Skills Analysis — 314 malicious skills from single user
Zen 🧘 (strategy, research, scoring design) & Neo ⚡ (PageRank implementation, CLI, A2A integration)
Built on A2A Secure v0.8 | February 2026
MIT