An agent trajectory is the ordered sequence of observations, reasoning and actions an AI agent takes while working on one task, from the first input to the final result. Researchers use the word for the same thing engineers call a trace: one complete run, usually stored as JSON with every message, tool call and observation.
Why it matters when licensing company data
Open formats now standardise trajectories so runs from different agent frameworks can be compared and reused. Harbor's Agent Trajectory Interchange Format records an agent's complete interaction history as JSON, including messages, reasoning, tool calls, observations and metrics, and its records carry the agent name, version and model.
Trajectories are used to fine-tune agents, to distil a stronger model into a smaller one, to build evaluations from real failures and to seed reinforcement learning environments. Each use needs the full sequence. A trajectory cut down to the first prompt and the last answer loses the steps where an agent re-reads, picks the wrong tool and recovers.
See licensing agent traces from production and RL environments built from company data.
Example
An engineering team exports a month of runs from its on-call agent and converts each one to a trajectory file: the alert text, forty ordered steps of log searches, metric queries and model reasoning, and the root cause the engineer on call accepted. Runs where the agent gave up are kept and labelled as failures. The team stores the model version and agent framework version in the first record of each file, so anyone reading the set later knows exactly what produced it.
Source
Harbor: Agent Trajectory Interchange Format (ATIF). Definition reviewed November 9, 2026 by Origin Data Partners.