Reputation & Trust Signals
How audit logs and agent activity feed into scoring, risk assessment, and policy enforcement
Reputation & Trust Signals
Key Takeaways
- Reputation scoring allows dynamic enforcement based on agent behavior - Trust signals can be derived from audit logs, revocation events, and verifier outcomes - Scoring models range from boolean allowlists to continuous trust scores - Verifiers can enforce risk-based policy at request time using reputation
Why Reputation Matters
In autonomous systems, the identity of an agent isn't enough. Reputation allows services to assess how trustworthy an agent is, based on its behavior over time.
Overview
MCP-I supports KnowThat Reputation Engine, a deterministic, transparent, and gaming-resistant reputation scoring system for AI agents. It solves the critical trust problem in AI agent interactions by providing mathematically rigorous scores that fairly evaluate both new and established agents with:
- Audit events emitted by agents or services
- Delegation chain evaluation
- Revocation history
- Optional external scoring registries (e.g., knowthat.ai)
This allows a verifier or relying party to adapt access controls, rate limits, or challenge levels dynamically.
System Overview
The KnowThat Reputation Engine uses a modular Rust workspace architecture:
knowthat-reputation-engine/
├── Cargo.toml (workspace)
└── crates/
├── reputation-types/ # Shared type definitions
├── reputation-core/ # Core algorithm implementation
├── reputation-wasm/ # WebAssembly bindings
└── reputation-tests/ # Integration test suite
Each crate has specific responsibilities with clear dependency boundaries:
-
reputation-types: Zero dependencies, shared across all crates
-
reputation-core: Depends only on types, implements algorithm
-
reputation-wasm: Bridges core to JavaScript world
-
reputation-tests: Tests the entire system
Scoring Models
Scoring is based on several factors:
- Conformance level
- Age of agent
- Number of past interactions
- Review scores and quantity
Integration Points
Reputation scoring pulls from logs generated by:
Scores can be:
- Stored locally
- Synced with a global registry
- Embedded in delegation metadata or verifier headers
Knowthat.ai Registry
KYA-powered deployments may optionally sync trust signals to a global registry at knowthat.ai, which supports:
- Agent lookup by DID
- Federation-safe trust flagging
- Shared deny/allow lists
This is optional and not required by the MCP-I spec.
Next Steps
→ Learn how audit logs are generated in Audit Layer
→ See how verifiers use trust signals in Edge Verification Guide