Glossary
Behavioral evaluation
A test that checks one observable agent action, such as a tool call or a question, rather than a whole benchmark score.
A behavioral evaluation (also called a behavioral eval) checks one action you can see in a run. The action might be a tool call, a clarifying question, a file edit, or a test command. The check reads that step. It does not, by itself, score whether a whole project succeeded.
Google’s September 9, 2026 post treats these checks as integration tests for an agent harness. An end-to-end benchmark still answers whether the job finished. The longer guide is How do you evaluate an AI agent harness?. Designing the system those tests wrap is harness engineering. A paper that grades entire harnesses on held-out tasks is HarnessFlow.