The full sequence of what an agent did: every thought, tool call, result and decision. It is the thing you debug, and it is the thing you evaluate, because an agent can reach the right answer through a process you would never sanction.
Technical view
The ordered record of states, actions and observations for a task episode. Used for debugging, for evaluation of process as well as outcome, and as training data for imitation or preference learning. Storage grows quickly, so retention policy matters, and trajectories often contain sensitive data from tool results, which brings them within data protection scope. Comparing trajectories across runs is how non-determinism in agent behaviour is quantified.
Related
Part of AI Terminology Index, section: Agents and Tools.