Trajectory-level monitoring is the practice of observing an AI system’s full sequence of actions over time, rather than judging each output in isolation. For agentic systems, a “trajectory” can include prompts, tool calls, state changes, network requests, retries, and the final result. This gives defenders a timeline of what the system actually did, not just what it said.
This matters because many AI security failures only become visible across multiple steps. A model may appear harmless in one response but later chain actions into data exfiltration, privilege misuse, policy evasion, or unsafe tool access. Trajectory-level monitoring helps detect abnormal paths, enforce guardrails, and reconstruct incidents for investigation. In practice, it relies on detailed logs, traceable tool execution, and alerting on risky behavior such as repeated failed attempts, unusual destination access, or unexpected escalation across an agent workflow.