What is an agent trajectory?

An agent trajectory is the full record of one run: every state the agent saw, every action it took, and every observation that came back, in order. The term comes from reinforcement learning, where a trajectory is the sequence of state-action pairs an agent produces across an episode. Not every agent benchmark grades the whole thing, though: Gaia2 checks each of the agent’s actions against an oracle sequence, step by step, while SWE-bench Pro and tau-bench only check the state the agent left behind, a patch that must pass a held-out test suite, or an account after a support call, and ignore how it got there. A trajectory is also what tool calling produces mechanically: each cycle of picking a tool, filling in its arguments, and reading back the result is one step, and the trajectory is that loop’s output, written down in order. Reading the trajectory differs from reading the final answer or final state: the answer tells you what the agent concluded, the trajectory tells you how it got there, which tool call was wrong, which observation it misread, or where a loop started repeating instead of making progress.