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

# Configure a reusable agent definition

> Select a bounded model, instructions and dependencies, then verify the published definition in Explorer.

You need Vantage, Forge authoring and access to the dependencies. Choose a test workspace and a representative question using your own permitted data.

## Configure a focused definition

1. Open **Forge → Agents** and create a definition for one clear purpose.
2. Write instructions describing the task, required evidence, uncertainty and stopping conditions.
3. Select a permitted model and supported runtime or harness requirements.
4. Add only the tools, groups and skills needed for that purpose.
5. Inspect required dependencies and any reason a capability is unavailable.

A selected tool grants no dataset access. A skill's requirements also grant no authority. The submitting person's permissions govern the actual turn.

Use the workspace's standing instructions for rules common to its agent chats. An Owner edits those separately under **Workspace settings → Agents → Instructions**. Workflow Agent steps use their block configuration and attachments.

## Publish within the supported scope

Inspect the draft, selected versions and audience. Publish it and make it available to the intended workspace.

V2 has no agent evaluation facility for deployment-wide sharing. Publishing or widening an agent for the whole deployment is refused. A copied agent starts private and unpublished.

## Verify in Explorer

Open the test workspace and start an agent chat using the published definition. Confirm its model and available capabilities.

Send your representative question. Check one retrieved source, one inference and any missing input. If you test delegation, confirm that child work remains within the parent task's scope and limits.

Try a request outside the definition's intended scope. Inspect the limitation rather than broadening permissions to make the test pass.

## Understand later changes

| Change | Effect |
| - | - |
| Publish revised instructions or dependencies | A later chat turn resolves the new published version. The current turn keeps its manifest. |
| Change the definition's model | An already-open agent chat keeps its model. Start a new chat to use the revised selection. |
| Edit workspace standing instructions | The next dispatched workspace-agent turn receives them. A message appended to an active run keeps that run's instructions. |
| Narrow or revoke a dependency | The next protected use must satisfy current access and version checks. |

Retain the test question, selected definition version and inspected result. Continue with [Investigate in Explorer](/vantage/explore-evidence/investigate).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.