Agent policy layer for safe tool use across MCP, A2A, and orchestrated agents.
Four tools with complete schemas and descriptions, but quality is uneven. All tools have input schemas with typed properties and descriptions, meeting baseline requirements. However, descriptions are terse (avg 65 chars vs production baseline 194 chars), lack WHEN/WHY guidance, and omit error recovery hints. Tool naming follows verb_noun convention (get, evaluate, evolve) but lacks clarity on distinctions, 'policy.evaluate' vs 'policy.evolve' could confuse LLMs about which to call first. No output schemas documented. Error handling is absent, no guidance on retryability, user-fixable vs fatal errors, or recovery paths. Security-critical tool (capabilities.get issues tokens) lacks permission/scope documentation.
Issue a capability token for an agent with specified subject context and requested policy packs
Retrieve published JWKS and policy pack metadata for token verification
Evaluate whether an action is allowed under a capability token and policy
Approve or deny a policy suggestion to evolve the policy state
No output schemas documented. LLMs cannot plan downstream calls or extract required fields (e.g., does capabilities.get return token expiry? does policy.evaluate return a decision reason?). Breaks tool chaining.
Descriptions are terse (40 - 55 chars) and lack WHEN/WHY context. 'Evaluate whether an action is allowed' does not explain when to call this vs capabilities.get, what happens on deny, or how to interpret the result. Production baseline is 194 chars.
No error handling guidance. If capabilities.get fails (invalid subject, unknown pack), LLM has no recovery path. No distinction between retryable (transient) and user-fixable (invalid pack name) errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | 2025-06-18+ | v2 |
Security-critical tool (capabilities.get issues capability tokens) lacks scope/permission documentation. No audit trail guidance. Violates pattern:scope-declaration and pattern:audit-trail.
policy.evolve accepts 'suggestion_id' but no prior tool returns suggestion IDs. Breaks tool chaining, LLM cannot discover how to obtain a suggestion_id without external knowledge.