Concepts

Agent trust score

A continuous, signable, offline-verifiable per-agent trust score — a deterministic blend of real signals, coverage-honest by design.

The Agent Trust Score is a continuous, per-agent score whose math you can verify. It is a deterministic (no-ML) blend of real signals:

• Signed authority — the agent's passport state.
• Delegation narrowing — whether each delegation hop actually narrows authority (a child can never hold more than its parent granted).
• MCP supply-chain integrity — whether the agent's tools still match their signed pin.
• Behavioral baseline — how far the agent has drifted from its own established pattern.

Coverage-honest by design

The score never invents a number to look complete. A signal with no data shows n/a — never a fabricated 0 — and an agent with no signals at all reads insufficient data rather than a made-up score. Because the blend is deterministic, you can open any score and see the weighted component breakdown and the hash of the inputs that produced it.

Signed and offline-verifiable

Each score can be signed into the Trust Ledger as a new point in the agent's score history, with per-point Merkle inclusion verification. Sign current score appends a verifiable point you can later check offline.

Where you see it

The ranked board lives on Agent Trust (/agent-trust, badged NEW), reading GET /api/agent-trust (and GET /api/agent-trust/[agentRef] for a single agent). Use it to spot weak agents and drill into why their score is what it is.