Agent Action Audit Record

Posture: docs_first, advisory_first. Describes a record’s required fields; implements no runtime, logger, or storage. Provider-agnostic.

Goal

Define the minimum record that proves what happened when an agent took (or was denied) an action: which skill, which task, which tool, what risk, which verdict, what evidence, and whether a human approved. “Audit proves” — without this record, a permission decision is just a claim.

Minimum record

agent_action_audit_record:
  task_id:
  agent_id:
  skill_id:
  autonomy_level:        # observe | advise | act_with_approval | autonomous_within_bounds
  tool_name:
  action_type:
  risk_class:            # coarse (low|medium|high|critical) or authoritative class
  policy_version:
  policy_verdict:        # allow | deny | require_approval | transform | log_only
  approval_required:
  approval_by:
  evidence_refs:
  knowledge_packs_used:
  restrictions_applied:
  timestamp:
  result_summary:

This is the minimum for declaration and review. It is not a new logging system.

Reconciliation with existing observability

The runtime harness already defines a per-event and per-action record:

• trace_event / minimum_events in agent-harness-governance-extension.md;
• the per-action JSON Schema agent-harness-governance-record.schema.json;
• knowledge-side logging in knowledge-audit-log-spec.md;
• slice-level telemetry in slice-telemetry-model.md.

This audit record is a skill-and-policy-oriented view over those: it names the fields a reviewer needs to connect skill → tool → verdict → evidence → approval. When emitting real traces, populate the governance record/schema; this document says which of those fields are non-negotiable for audit of a skill-driven action. Where they overlap, the governance record schema is authoritative for structure.

Non-negotiable fields

At minimum, a recorded action must carry: skill_id, task_id, tool_name, policy_verdict, evidence_refs, and — when the verdict was require_approval — approval_by.

See also

• policy-verdicts.md — the policy_verdict values.
• agent-harness-governance-extension.md — trace events + record schema.
• autonomy-levels.md, tool-risk-taxonomy.md.