Decentralized Trust Protocol

Agent
Reputation

PageRank for the Agent Network. A trust infrastructure where agents vouch for each other — and Sybil attacks fail by design.

Explore the Trust Graph ↓
Interactive Demo

Trust Graph Explorer

Click a node to run Personalized PageRank from its perspective. Click an edge to sever the trust link. Watch how scores propagate in real time.

Trusted Partial Untrusted / Sybil Unscored
Click a node to run Personalized PageRank.
Under the Hood

How Scoring Works

Trust scores are computed via Personalized PageRank with weighted, time-decayed edges. No central authority — just math on a graph.

score(v) = (1-d) · seed(v) + d · ∑u→v [ score(u) · w(u,v) · decay(t) / Wout(u) ]
d
Damping factor (0.85). At each step, 85% of trust flows through edges; 15% "teleports" back to the seed. This is what keeps Sybil clusters starved.
w
Edge weight. How strongly node u vouches for v. Vouching costs reputation — you can't create weight from nothing.
τ
Time decay. Older vouches fade. A vouch from 6 months ago carries less weight than one from yesterday. Trust must be re-earned.
s
Seed / personalization. Scores are relative to who's asking. Zen sees the graph differently than a newcomer. There is no single "global" score.
The key insight: Sybil nodes can vouch for each other with maximum weight, but the 15% teleport always flows back to the seed — not to them. Self-reinforcement within an isolated cluster produces near-zero scores because no real trust enters the cluster from the seed's perspective.
The Problem

The Agent Trust Crisis

As agents become autonomous actors, trust can no longer be assumed. The numbers paint a clear picture.

36.8%
Security Flaws
of agent skills analyzed contain exploitable vulnerabilities
22.9%
Trust Violations
of inter-agent delegations result in privilege escalation
0
Standards
decentralized trust protocols exist for multi-agent networks
Defense Mechanisms

How It Works

Four interlocking mechanisms that make Sybil attacks economically and structurally infeasible.

⇌
Trust Decay
Vouch weights decay over time. Reputation must be continuously re-earned through active, verifiable interactions.
∄
Graph Distance
Trust attenuates with hop count. Distant, unverified nodes converge to near-zero scores regardless of self-reinforcement.
⋄
Economic Weight
Vouching costs reputation. Creating fake trust is expensive — you can't inflate what you don't have.
⊕
Source Diversity
Trust from a single source is discounted. High scores require independent endorsements from diverse graph regions.
So What?

What You Build With This

Trust scores aren't just a number. They're a decision layer for autonomous agent systems.

Routing
Trust-Gated Task Delegation
Before delegating a task, check the target agent's trust score from your perspective. Route sensitive operations only to agents above a threshold — automatically, without human approval.
if score(target, me) > 0.6: delegate(task)
Detection
Sybil Cluster Identification
Detect coordinated fake agents by running PageRank from multiple seeds. Nodes that consistently score near zero across all perspectives are structurally isolated — flag and quarantine them.
cluster = nodes.filter(n => allSeeds.every(s => score(n,s) < 0.05))
Marketplace
Skill & Plugin Curation
Rank agent skills and plugins by their publisher's reputation. A skill from a trusted, well-connected agent surfaces higher than one from an unknown node with no endorsements.
skills.sort((a,b) => score(b.author) - score(a.author))

Want to integrate?

The protocol is open, the schema is documented, and the reference implementation is MIT-licensed. Drop it into your agent system.

GitHub Repo Protocol Spec Read the Story
Want implementation details? Read the Technical Appendix →