agentmesh-governance_skill

Overview

This skill provides a zero-trust governance layer for AI agents using AgentMesh to enforce policies, compute trust scores, verify cryptographic identities, and maintain tamper-evident audit logs. It is designed for command-line integration so teams can gate tool calls, enforce token and tool limits, and audit agent behavior before, during, and after execution. Enterprise-grade testing and broad adoption make it suitable for production safety and compliance workflows.

How this skill works

The skill evaluates proposed agent actions against a YAML policy to allow or deny execution, returning JSON with verdicts, violations, and recommendations. It produces a composite trust score across five dimensions (policy compliance, resource efficiency, output quality, security posture, collaboration health) and updates scores after interactions. Identity verification uses Ed25519 DIDs and signatures, while audit trails are recorded in a Merkle chain so integrity checks detect tampering.

When to use it

• Before executing any tool call to enforce token limits, allowed tools, and content policies.
• When deciding whether to delegate work or collaborate with another agent by checking its trust score.
• To verify an agent's identity and signed messages using Ed25519 DIDs before trusting outputs.
• After agent collaboration to record outcomes and update trust scores.
• For compliance or incident investigations by verifying Merkle chain audit logs.

Best practices

• Integrate policy checks into the pre-execution path so denials happen before any external call.
• Define conservative default policies (lower token budgets, strict allowlists) and relax per-environment as needed.
• Require human approval for high-risk actions via policy flags instead of ad hoc decisions.
• Regularly review trust score weightings and retrain any heuristics driving the output-quality dimension.
• Rotate keys and reissue DIDs on a schedule; verify signatures on every cross-agent communication.

Example use cases

• Prevent runaway token consumption by enforcing per-action and per-session token caps for generative tools.
• Block dangerous tools (shell_exec, file_delete) and destructive content patterns before execution.
• Verify another agent’s DID and signature before accepting a generated data feed or document.
• Maintain an auditable, tamper-evident trail of agent actions for compliance and post-incident review.
• Automatically block or quarantine agents whose trust scores fall below a configured threshold.