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

# Eval Health Agent

> Reviews the evals running on a tracing project, finds missing coverage and evals that are wasteful or uninformative, and proposes fixes you approve in one click.

> [!NOTE]
>
> The Eval Health Agent is part of managed agents on **Enterprise**. Signal and managed agents are currently included at no additional charge for Enterprise. See [Pricing and usage](/docs/ax/security-and-settings/pricing-and-usage.md).

The **Eval Health Agent** reviews the evals already running on a tracing project, works out which ones the app is missing, and flags evals that cost more than they need to or no longer tell you anything. Each finding is filed as an issue with a proposed change attached. **The agent never creates or edits an eval on its own**: nothing changes until you accept the proposal.

## What it finds

| Finding | What it proposes |
| - | - |
| **Coverage gap**: a quality dimension your app needs but no eval measures | A new evaluator and a continuous eval task on the project |
| **Broken or stale eval**: wrong variable mappings, a template that no longer fits the traces | A fixed evaluator version or task configuration |
| **Inefficient eval**: an LLM judge doing work a code eval or cheaper model could do | A code eval, or a change to the judge model |
| **Ineffective eval**: an eval that is not providing meaningful signal | Removing the eval from the task |

Each run proposes at most **5** new evals. Optimizations and fixes don't count toward that cap. Most runs reconfirm what is already filed, so a run that files nothing new is a valid result.

## Run it

**Signal & Agents → + New Agent → Monitor & optimize → Eval Health**, or open a tracing project and choose **Agents → Eval Health → Enable Agent**.

| Setting | Default |
| - | - |
| **Project** | Required. One Eval Health agent per project. |
| **Trigger** | Every 6 hours. Editable before you save. |
| **Agent** | Claude Code on the Arize-managed provider. You can switch harness or use your own AI provider. |
| **Name** | `Eval Health — <project>` |

> [!NOTE]
>
> **It needs a space default LLM integration with a default model.** The evals it proposes are LLM-as-a-judge evaluators that run on that integration. If the space has none, setup stops with a prompt to set one in **Space Settings**. See [AI provider integrations](/docs/ax/security-and-settings/integrations-playground/overview.md).

The first run starts as soon as you save, then it follows the schedule. Pause or resume it from the settings menu on the **Eval Health** tab.

## What it reads

Each run reads the project's eval tasks, the space's evaluators, and up to **100 spans from the last 7 days** to see what the app actually does. Before proposing a variable mapping, it checks the mapping resolves against a real span.

## Review findings

Findings appear on the project's **Eval Health** tab, which shows once the project has an agent, and on the space-wide **Issues** page with an **Eval Health** badge. The tab summarizes healthy evaluators, coverage gaps, efficiency and effectiveness findings, and how many proposals need review.

Each issue carries an action card. Accepting runs with **your** permissions, and every card also has **Dismiss**.

| Card | Button | What accepting does |
| - | - | - |
| **Create eval** | **Accept and run** | Creates the evaluator and a continuous eval task on the project. Turn it off any time in Eval Hub. |
| **Update eval** | **Apply fix** | Saves a new evaluator version and/or updates the task. Earlier versions stay in Eval Hub. |
| **Optimize eval** | **Apply** | Changes the judge model, or replaces the LLM judge with a code eval. Nothing is deleted. |
| **Remove eval** | **Remove** | Stops this task running the evaluator. The evaluator stays in Eval Hub. |

> [!TIP]
>
> If the eval changed after the agent proposed a fix, **Apply fix** and **Apply** are rejected with "Eval changed since this fix was proposed" rather than overwriting your edit. Wait for the next run to propose against the current version.

If you lack permission to accept a card, it asks you to have a space admin accept it.

## Related docs

* [Managed agents](/docs/ax/agents/build-your-own-agent.md): the gallery, agent configs, and customization
* [Dataset Coverage agent](/docs/ax/agents/dataset-coverage-agent.md): keep a dataset in step with production traces
* [Cost agent](/docs/ax/agents/cost-agent.md): ranked cost issues with quantified savings
* [Signal](/docs/ax/observe/signal.md): ranked correctness and reliability issues