Agent Evaluation Harness

An agent without an eval harness is a demo, not a system.

You cannot ship, refactor, or compare agents responsibly without one. Every prompt tweak that "feels better" is a coin flip until you have a number to back it.

When to use this skill

• The user is about to merge a prompt or graph change.
• The user is choosing between two models, two patterns, or two providers.
• The user is debugging a regression and has no way to reproduce it.
• The user is asking "is this good enough to ship?"
• The user mentions LLM-as-judge, golden set, eval suite, or regression test.

What an eval harness actually is

Three pieces. If any is missing, you don't have one.