Bottom Line: Agent evaluations are more complex than traditional LLM tests because they involve multiple turns, tool usage, and state changes. The key is distinguishing between transcript (recorded interactions) and outcome (actual final state) to create meaningful assessments.
Good evaluations enable teams to deploy AI agents more reliably. Without them, developers easily fall into reactive loops and catch errors only in production. This guide demonstrates best practices for rigorous agent evaluations.
Evaluations (“evals”) are tests for AI systems: you provide an AI with an input and apply evaluation logic to the output to measure success. While single-turn evaluations are relatively straightforward—a prompt, a response, and evaluation logic—multi-turn evaluations are becoming increasingly common with advanced AI models.
With agent evaluations, it becomes significantly more complex. Agents use tools across many turns, modify the environment’s state, and adapt—which means errors can propagate and amplify. Frontier models can also find creative solutions that go beyond static evals.
When building agent evaluations, the following definitions are central:
A **Task** is a single test with defined inputs and success criteria. Each attempt to execute a task is a **Trial**. Since model outputs vary between runs, you conduct multiple trials.
A **Grader** is logic that evaluates an aspect of agent performance. A task can have multiple graders, each with multiple assertions.
A **Transcript** (or Trace/Trajectory) is the complete dataset of a trial—including outputs, tool calls, reasoning, intermediate results, and other interactions.
The **Outcome** is the final state of the environment at the end of the trial. A flight booking agent might say “Your flight has been booked,” but the outcome is whether a reservation actually exists in the environment’s SQL database.
An **Evaluation Harness** is the infrastructure that executes evals end-to-end, automating the execution, measurement, and analysis of tests.
Source: www.anthropic.com