Agent trajectory
An agent trajectory is the complete recorded path of one agent run: the sequence of observations, reasoning steps, actions, and resulting states from task start to termination. Trajectories are the unit of analysis for agent work, the way sessions are for human users; you evaluate, compare, and replay trajectories, not isolated actions.
How it works
A trajectory interleaves data from different sources: model inputs and outputs, tool calls with arguments and results, and, for browser agents, the evolving interface state. The interface portion is where recording format matters: a DOM event stream captures each state compactly and diffably, so two trajectories of the same task can be compared step by step ("run A saw the modal, run B did not"), which screenshots make expensive and lossy. Stored trajectories accumulate into a corpus that serves three consumers: debugging (replay the failure), evaluation (score outcomes across runs), and training data.
Where you encounter it
Agent eval harnesses, regression analysis between model or prompt versions, and replay tooling for agent trajectories.