Getting an agent to work once is not enough. Agents are sensitive to prompts, model versions, retrieved context, tool output, latency, rate limits, and small changes in user input. Evaluation is how you find out whether the system works consistently, not just whether it had one good run.

Agent evaluation asks a practical question: does this system complete the task correctly, safely, and efficiently under the conditions it will face in production?

Answering that takes more than a few manual prompts. You need a repeatable set of test cases, stable tool responses where possible, clear scoring rules, cost and latency tracking, and regression tests that run before changes ship.

This chapter shows how to build that kind of evaluation without turning it into a research project.