Agents and tool evaluation

How do you architect an evaluation pipeline for autonomous agents, specifically measuring tool execution and reasoning trajectories?