Skip to content
Outrich Data

Glossary

Agent trajectory

An agent trajectory is a recorded sequence of the steps taken to complete a task: what was observed at each point, what action was taken, which tool was used and what resulted. AI labs use trajectories to train and test agents, meaning models that carry out multi-step work across software systems.

An answer shows a result. A trajectory shows the route. For an agent, the route is what has to be learned: open the record, check the account status, look up the policy, issue the credit, write the note, close the case. Each step has a state before it, an action and a state after it. A complete trajectory ends with an outcome that tells the model whether the sequence worked.

Companies hold trajectories without calling them that. Workflow systems log state changes and approvals. Ticketing tools record every action on a case. Audit logs capture which screens and records a user touched. Version control shows the order of edits to a code base. The most useful trajectories link these sources through a shared case identifier, so a task can be followed across tools. Trajectories that include errors and corrections are valuable, since they show recovery. Preparing them means preserving order and timestamps, hashing identifiers consistently so the links survive, and removing personal details from the content of each step.

Trajectories are usually delivered as JSONL, one task per line, with the steps held in an ordered list inside each record.

Example

A refund case: the agent views the order, checks the return policy, confirms delivery status, approves a partial refund and records the reason. Five ordered steps and a final outcome form one trajectory.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc