interactive · 12 min
An Eval Harness
Objective: run an agent over a labelled set of cases and compute a pass rate — the smallest useful evaluation.
MockLLM is the same Foundations seam; here it’s the system under test.
The harness is just: run every case, compare to the label, aggregate.
Everything sophisticated (judges, trajectory checks) is a richer
comparator wrapped around this same loop.
The same loop measures retrieval recall@k: the case label is whether the
expected document appears in the top-k retrieved, and the comparator checks
membership instead of equality.
Best practice: start every agent with a tiny golden set and a pass rate in CI; grow the set from every production failure.
Next: Evaluation Pitfalls