How we know whether the behavioral layer actually holds. This section covers evals, benchmarks, observability, and red-teaming, with an emphasis on behavioral measurement (does the agent escalate, disclose, and recover correctly?) rather than raw capability scores alone.

What belongs here

  • Benchmarks and eval suites for agent behavior, tool use, and task completion.
  • Observability and tracing approaches for understanding agent decisions in production.
  • Red-teaming and adversarial testing aimed at behavioral failure modes.
  • Scoring methods, rubrics, and judge designs for open-ended agent output.

What goes elsewhere

  • The behaviors being measured are defined in Behavior.
  • The harnesses and frameworks under test live in Build.

See tau-bench for a worked example in this section.