A tool call is a request an AI model makes, during an agent run, for the harness to run a named function with specific arguments, such as searching logs, reading a record or sending a message. The harness runs it and returns the result to the model. Both the arguments and the result are part of the agent trace.
Why it matters when licensing company data
In the OpenTelemetry GenAI conventions each tool call is an execute_tool span, with the parameters passed to the tool and the result it returned held in opt-in content attributes. Tracing tools show the same pair as a tool span under the model call that requested it.
Tool results are usually the largest and most sensitive part of a trace, because they copy in whatever the tool read: database rows, ticket threads, files, log lines. They are also where an export is most often cut short. A trace that keeps the call but drops or truncates the result no longer shows what the agent saw when it made its next decision.
See how to export agent traces and the agent trace redaction checklist.
Example
An IT help-desk agent receives a request to restore a user's access. It makes three tool calls: a directory lookup with the user's ID, a query of recent access changes and a request to re-enable the account. The trace records each call's arguments and the full result returned. Before the team shares the run internally for review, it replaces the user's name and ID with consistent placeholders in both the arguments and the results, so the run still reads end to end.
Source
OpenTelemetry GenAI semantic conventions: execute_tool span. Definition reviewed November 9, 2026 by Origin Data Partners.