Evaluation (evals)
Systematic tests that score an agent's outputs against expected results or a rubric.
Offline evals run the agent on a fixed test set before release; online evals score live traffic. Scoring is by exact match, execution (do the tests pass?), human review or LLM-as-judge. LangChain's 2026 report: 52.4% of teams run offline evals, 37.3% online, 29.5% none; among evaluating teams 59.8% use human review and 53.3% LLM-as-judge.