Getting an agent to work once is not evidence that it works. Agent behavior changes with prompts, model versions, retrieved context, tool outputs, latency, rate limits, and small differences in user input. Evaluation is how you turn a promising demo into an engineering artifact you can reason about.

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 requires more than a few manual prompts. You need versioned eval datasets, deterministic tool fixtures where possible, measurable outcome metrics, cost and latency tracking, trajectory review, and regression tests that run before changes ship.

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