If you don't have an eval harness, you don't have a production AI system.
Evals Are the New Tests Traditional software has unit tests. AI systems need evals — graded assessments of model behavior on representative inputs. Three Types 1. Reference based : Compare output to ground truth. Good for classification, extraction. 2. LLM as judge : Score outputs with another LLM. Good for quality, tone, helpfulness. 3. Behavioral : Test specific scenarios end to end. Good for agents. Without evals, you're prompt tuning blind.