A buying reference for agent training data

Content reviewed

Best AI Agent Training Data Procurement Reference

Glossary Tool-use trajectory

Tool-use trajectory

The ordered record of one agent run: which tool was called, with which arguments, what came back, and where the run stopped.

Answer in brief

A tool-use trajectory is the ordered record of a single agent run: each tool call, its arguments, the observation the tool returned, and the reason the run stopped. It lets a reviewer follow a decision without rerunning the task.

Trajectories are not transcripts. A log that omits arguments or tool output shows that a tool was used, but not what the agent knew when it chose the next step — the part a policy has to learn.

Definition

A tool-use trajectory is the ordered record of tool calls, arguments, observations and outcome across one task attempt. Order carries meaning: the same calls in a different sequence usually represent a different policy, and sometimes a failed run that ended in the right state.

Trajectories are bought separately from demonstrations because they expose intermediate decisions. A demonstration is judged on its accepted action; a trajectory has to hold up step by step, including the steps that went nowhere.

Minimum schema

Agree these fields before collection starts:

  • A step index, so two reviewers can refer to the same step.
  • The tool name and the arguments exactly as sent, not a summary of them.
  • The observation returned, truncated only where the record says so and points to the full value.
  • The stop reason: task complete, budget exhausted, blocked, refused, or escalated to a person.
  • The final state or output that the success judgment refers to.

Forbidden tools and stop conditions

List the tools that must never be called and the data that must not leave the environment, then check those rules against the runs you are shown. The sandbox environment entry covers how permitted actions are bounded.

Stop conditions belong in the same specification: a step budget, a time limit, and a requirement to escalate when a permission check fails. A run that passes its stop condition is a defect even when the final answer looks correct.

Partial credit versus final success

Keep two judgments apart: whether the final state met the objective, and whether the steps were acceptable. A run that reaches the right state by calling a forbidden tool has failed; a run that fails at the last step may still contain a correct plan worth learning from.

Partial credit is useful for training signal and risky as an acceptance metric unless the weights were agreed in advance. Ask for the scoring rule, then apply it to runs you have already read. The evaluation set entry explains why the scorer version has to travel with a score.

Acceptance question

Request complete runs for one task of your own, including a failed run. Check that the record reads without the supplier's tooling, that observations are present rather than referenced, and that the stop reason matches what you can see. The tool-use guide sets out the fields in order; the buying guide places this check inside a pilot.

Questions & answers

Is a chat transcript a tool-use trajectory?
No, unless it carries the tool calls, exact arguments, observations and stop reason in order. A transcript usually shows that a tool was used while hiding what it returned, which removes the information that explains the next decision.
Should failed runs be delivered?
Yes, labeled with their stop reason and not mixed silently into accepted runs. Failures show where the policy breaks; ask whether the label separates a wrong plan from a tool that refused to run.

Sources for this entry

  1. Editorial checklists on this site — Original guidance written for this reference; the schema fields are editorial recommendations, not a published standard.
  2. GAIA: a benchmark for general AI assistants — Public paper describing real-world assistant questions that require tool use; cited for the category, not for any supplier claim.

Editorial checklists on this site are original guidance. Links to official product documentation help readers understand the category; they do not establish comparative performance and are not endorsements of any supplier.

Cite this entry

Best AI Agent Training Data. “Tool-use trajectory.” Content reviewed .

https://bestaiagenttrainingdata.com/