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

# Configure a Custom Agent

> Choose settings, tools, skills, files, judges, target accounts, memory, and schedule.

Open an agent and select **Configure**. The available configuration areas can include **Settings**, **Tools**, **Skills**, **Files**, **Judges**, and **Versions**.

## Settings

Configure the model, status, labels, output, scope, and schedule.

### Targeting

* **Organization** runs once with organization-level scope.
* **Accounts** runs for selected accounts.
* **Dynamic** targeting uses the configured criteria to choose accounts at run time.

For account-targeted structured output, **Gather Results** can combine per-account results into a collection. Per-account schedules and context may also be available.

### Schedule

Schedules use the organization's configured timezone. Document the expected completion window and the person responsible for late or failed runs.

### Structured output

Use a structured output when another process or reviewer needs consistent fields. Include a narrative explanation when the fields alone would hide important uncertainty.

### Shared memory

Shared memory preserves agent-specific context across eligible runs. It is off by default. Enable it only when prior-run knowledge is necessary and appropriate for the targets.

<Note>
  Dry runs may read existing shared memory but do not add new memory. Review memory behavior before using it for account-sensitive decisions.
</Note>

## Tools

Select only the tools required by the workflow. The catalog already reflects system connections, organization policy, and your own access. A selected tool must also be available to the Custom Agents surface when the agent runs.

## Skills and files

Attach reusable skills when the workflow should follow maintained guidance. Attach stable reference files, not transient evidence that belongs to one run.

## Judges

Judges evaluate run quality or adherence to a defined criterion. Create judges after the agent has been saved. A judge should be specific enough that a failed score suggests a concrete response.

Examples:

* required evidence is cited;
* all configured accounts are represented;
* output contains no unsupported recommendation;
* escalation is present when a stop condition occurs.

## Versions

Use Versions to compare meaningful changes and understand which configuration produced a run. Test after changing instructions, tools, skills, output structure, targeting, memory, or schedule.