> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cognite.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Running agent evaluations

> Test Atlas AI agents to verify that they provide the results that you want and identify where they can improve performance.

Configure evaluations to verify that your agents provide the results that you expect. Create test cases with prompts and expected responses, run evaluations to generate agent responses, then review results to identify where your agents can improve.

## Before you start

You must create an agent in the [Agent builder](/cdf/atlas_ai/guides/atlas_ai_agent_building).

## Configure evaluation

Configure an evaluation with test cases that define prompts and expected responses.

<Steps>
  <Step title="Start configuration">
    1. In CDF, navigate to **Atlas AI** > **Evaluate agents**.
    2. Select **+ Create new evaluation**.
  </Step>

  <Step title="Enter evaluation details">
    1. Enter a **Name** for your evaluation.

    <Tip>
      Use descriptive names like "Pump maintenance agent - v1" to help you identify your evaluations.
    </Tip>

    2. Add a **Description** to explain the purpose of the evaluation.
  </Step>

  <Step title="Define test cases">
    Add test cases to define prompts and the expected responses.

    1. In the **Test cases** section, enter a **Prompt**.
    2. Enter the **Expected response**. Atlas AI compares your expected response to the agent's response to test performance.
    3. Optionally, choose an agent and select **Generate answer** to use the agent to generate the expected response.

    <Tip>
      Be specific about what details, like operational status, metrics, or alert information, you want agents to include in their responses.
    </Tip>
  </Step>
</Steps>

## Run and monitor evaluation

Run an evaluation to test your agent. Monitor real-time progress as the agent processes test cases.

<Steps>
  <Step title="Navigate to evaluation">
    1. Go to **Evaluation overview**.
    2. Select **Run evaluation**.
  </Step>

  <Step title="Select your agent">
    1. Use the search or filter to find your agent by name, status, or latest version.
    2. Select the agent from the list.
    3. Select **Run with selected agent** to run the evaluation.
  </Step>

  <Step title="Confirm code tool access">
    If Atlas AI prompts you to allow code tool access for your agent, select **Confirm and run** to continue running the evaluation or **Cancel** to go back to the list of agents.
  </Step>

  <Step title="Monitor evaluation progress">
    When you run the evaluation, Atlas AI sends your prompts to the agent and records the responses in real time. This can take several minutes depending on how many test cases you have.

    You can select **Cancel run** to stop the evaluation at any time.

    <Warning>
      Do not close the browser tab while the evaluation is running. Evaluations run in the browser, and closing or navigating away from the tab will stop the evaluation.
    </Warning>
  </Step>
</Steps>

## View and analyze results

Review evaluation results to identify where agents need improvement and track performance across multiple runs.

<Steps>
  <Step title="View completed evaluations">
    On the **Evaluation overview** tab, review the list of completed evaluations.
  </Step>

  <Step title="Review test case status">
    Select an evaluation to view the status of your test cases.
  </Step>

  <Step title="Compare responses">
    Select **View details** to compare the agent responses to your expected responses for each test case.

    <Tip>
      Use evaluation insights to improve your agent by refining instructions in the Agent builder, updating tools or data access, or editing test cases.
    </Tip>
  </Step>
</Steps>

## Related information

* [Building agents](/cdf/atlas_ai/guides/atlas_ai_agent_building)
* [About AI agent evaluations](/cdf/atlas_ai/concepts/atlas_ai_agent_evaluation_overview)
* [Agent tools library](/cdf/atlas_ai/references/atlas_ai_agent_tools)