Tracing is the security and observability record of an AI agent’s behavior: the steps it took, the tools it called, the inputs it used, and the state changes it caused. Unlike a simple chat transcript, tracing captures execution details that let defenders reconstruct what the agent actually did, not just what it said it would do.
Tracing matters because autonomous agents can read untrusted content, make decisions, and act across multiple systems. If an attacker uses prompt injection or manipulates an integration, traces help security teams see whether the agent merely encountered malicious data, attempted a risky operation, or changed a resource. Good traces support review, incident response, policy enforcement, and rollback. In practice, tracing is strongest when it is tamper-resistant, tied to identity and authorization events, and detailed enough to show every tool call and permission check.