agent-governancelisted

# Agent Governance ## Purpose and boundary Put enforceable policy between an Agent request and tool execution: ```text Request → Intent/Risk Check → Policy Decision → Approval (if required) → Tool → Audit ↓ ↓ ↓ Threat signal allow/deny/review Trust update ``` Governance is a control plane, not a prompt suggestion. Policies must be machine-enforced, versioned, reviewable, and fail closed when the decision, authority, or audit record is missing. Default to read-only discovery and a policy/report draft. Do not execute tools, expose sensitive content, grant permissions, or alter production policy without explicit authority. ## Governance contract Define before enabling a governed Agent: - policy identity, version, owner, scope, environment, and expiry/review date; - allowed tools and capabilities, blocked tools, resource/tenant scope, and data classes; - actions that require human approval, quorum, separation of duties, or re-authorization; - maximum calls, concurrency, runtime, tokens, network, storage, and spend; - content and intent signals, confidence threshold, escalation behavior, and appeal path; - audit events, retention, redaction, integrity protection, and authorized readers; - failure behavior, rollback/compensation, and the policy-change test plan. Separate **capability** (what a tool can do), **authority** (what this Agent may do now), and **approval** (who authorized this