What audit trail do I need to log for AI agent actions and tool calls?

For AI agent actions and tool calls, a forensically complete, immutable, and unified audit trail is required to reconstruct the full sequence of events, including what the agent saw, decided, and executed. This audit trail should capture decision context and be queryable for incident response, compliance, and internal review.

Concrete controls for logging AI agent actions and tool calls include:

• Unified Audit Trail (NIST AI RMF: Govern): Implement a unified governance layer that normalizes heterogeneous tool calls from different agent runtimes into canonical actions (e.g., read, write, execute) and records all agent actions in a consistent format, such as JSONL. This helps overcome "Audit Blind Spots" and "Compliance Complexity" by linking operations across agents and meeting compliance requirements like HIPAA, SOC 2, and PCI-DSS.
• Detailed Action Logging (NIST AI RMF: Govern): Log every action with details such as timestamp, agent identity, the specific action performed, the resource accessed, the environment, and the policy decision (e.g., PERMIT or DENY) along with the reason for denial. This level of detail is crucial for understanding "why it happened" and supporting accountability.
• Tool Call Logging (NIST AI RMF: Govern): For each tool call, record the tool name, its arguments, the result (or a preview of it), and whether the action was approved, especially for destructive actions. This helps in reconstructing what tools an agent called and the results it received.
• Immutability and Forensic Readiness (NIST AI RMF: Govern): Ensure that audit logs are immutable and append-only, preserving decision context and allowing for the reconstruction of events without alteration. This capability, often referred to as an "AI Agent Flight Recorder," is essential for determining the "blast radius" of a compromise and providing board-level accountability.
• Session Inspection (NIST AI RMF: Govern): Capture and allow inspection of the state of an agent run, including session ID, agent, status (e.g., success, denied), and timestamp. This provides a high-level overview of agent activity and outcomes.
• Contextual Logging (NIST AI RMF: Govern): Log not just what happened, but also the context, decision logic, and policy constraints that were in effect at the time of the action. This is vital for systems with growing autonomy and non-deterministic behavior, where understanding the "why" is as important as the "what".