How do you evaluate an AI agent's performance across a...

Agent evaluation needs several dimensions scored independently rather than one pass/fail number, because success and process quality can diverge in either direction:

• Task success rate — did it actually complete the task correctly?
• Step efficiency — how many tool calls did it take relative to the minimum needed?
• Tool selection accuracy — did it pick the right tool at each step?
• Reasoning quality — were the intermediate thinking steps sound, independent of whether the final answer happened to be right?
• Latency and cost — total time and total tokens or API spend for the task.
• Error recovery — when something went wrong mid-task, did it self-correct?

A “find and summarize the top 3 AI papers this week” task might come back with a correct 3-for-3 result — full marks on task success — while having used five tool calls when three would have sufficed, putting step efficiency at a middling score. A single success/fail metric would call that task a clean win and miss the inefficiency entirely; scoring each dimension separately is what surfaces it. The practical value of this scorecard approach is catching the gap between an agent that’s technically correct and one that’s actually production-ready.