Agent tool use extends an agent beyond generating a direct response. The agent can choose an available tool, provide the required inputs, interpret the result, and decide how that result affects the next step.
Extended tool use becomes harder as the workflow grows because small errors can affect later actions. Evaluation should consider whether the agent chose appropriate tools, maintained task state, verified outcomes, and recovered accurately from failures.
ELI5
Agent tool use is like an AI helper choosing the right item from a toolbox. It must know which tool fits the job and what to do with the result.
For example, the helper might use a browser to gather facts, a spreadsheet to organize them, and a drawing tool to make a chart, checking each result before moving on.
