Evaluating a Co-Worker
======================

Evaluations let you grade a Co-Worker's answers against expected results using a labelled dataset. Because a Co-Worker runs as an agentic flow, you evaluate it with an **In-House DeepEval** evaluation whose inference source is the agent itself.

.. image:: ../images/coworker-evaluation.png
   :alt: In-House DeepEval form with Inference Source set to Agentic Flow Agent and a Co-Worker selected
   :align: center
   :width: 100%

|

#. In the sidebar, open **Evaluations**, then the **In-House** tab. (If you do not see it, ask your administrator.)
#. Click **+** and choose **DeepEval**.
#. **Basic Information** — enter a Name (lowercase letters, numbers, and hyphens; up to 50 characters) and optional Tags.
#. **Metrics** — tick one or more; at least one is required.
#. **Inference Source** — choose **Agentic Flow Agent**.
#. **Select Agent** — pick your Co-Worker. Items marked *(Draft only)* have no published version; toggle **Use Draft Version** to test the unpublished draft.
#. **Evaluation Model** — choose the judge model (Model Category, then Model Type) that grades the answers.
#. **Dataset** — select a dataset with an ``input`` column and an ``expected_output`` column (the ground truth). If you have none, create one first.
#. Click **Save**, open the evaluation, go to the **Runs** tab, and click **Run Evaluation**. Use **Refresh** to update the status.
#. Open the completed run to review the **Overall Metrics** card and the per-row table — Input, Expected Output, and each metric's result with an explanation. Use **Download** to export the results.

.. note::

   Evaluations, the full metric catalogue, and datasets are documented under :doc:`Foundry → Evaluations </12_foundry/evaluations/index>` and :doc:`Foundry → Datasets </12_foundry/datasets/index>`. This page covers only the Co-Worker-specific path — an In-House DeepEval evaluation with the **Agentic Flow Agent** inference source pointed at your Co-Worker. To grade a model directly rather than an agent, use the Bedrock evaluation family instead.

.. tip::

   To build an evaluation dataset from real traffic, open a trace on the :doc:`Traces tab <observability>` and use **Add to Dataset**.