Draft an Agent with the Admin Assistant

Describe the job you want an Agent to perform. The Admin Assistant turns that description into a saved draft, checks it against the capabilities available in your Workspace, and leaves publication to you.

Goal

Finish with one Agent draft whose instructions, model, tools, resources, and permission boundaries you have reviewed. The Agent becomes available to new Sessions only after you explicitly publish it.

Prerequisites

• Run awaken and open the Console for the target Workspace.
• Configure at least one runnable provider-backed model. If Models shows no runnable choice, complete provider and model setup first.
• Decide the Agent ID, its first task, the tools it may use, the actions that require approval, and what it must never do.

1. Open the Assistant

Open /w/<workspace>/assistant in the Console. The Console prepares the reserved __admin_assistant Agent from a runnable Workspace model. It creates the conversation Session only when you send the first message, so opening the page alone does not leave an empty Session behind.

The Assistant can inspect the published capabilities in this Workspace. It does not read stored secrets, create credentials, or publish an Agent.

2. Describe the job and its limits

Give the Assistant a concrete first task and a finish line. Replace the bracketed tool IDs with tools that already exist in your Workspace:

Create an unpublished Agent with display name Support Triage and stable ID support-triage.

Its job is to read a support report, summarize the problem, and list the facts
still needed before an engineer can act. Use only [tool-id-1] and [tool-id-2].
Require approval before any tool action that changes external state. Never send
messages or change customer data. Prefer a published runnable model. Save and
validate the draft, but do not publish it.

Include examples when tone or output shape matters. If the request depends on a model, tool, plugin, Skill, MCP server, or Resource that is not available, choose an existing capability or configure it before continuing. Do not ask the Assistant to invent an ID that you will later have to interpret.

3. Review the saved draft

When the draft card appears, choose Open in editor. Review the configuration against the job you described:

• Do the instructions define the result and the stopping condition?
• Are the selected model and input modalities suitable for the first task?
• Is every tool, plugin, Skill, MCP server, and Resource necessary?
• Do permission rules stop the actions that require a person?
• Does the diff contain only the change you intended?

A clean validation result means the configuration can be published. It does not mean the Agent’s behavior is good enough for your users.

4. Check and publish the draft

Choose Check draft after any manual edit. Then choose Review & publish, read the exact draft-versus-published diff, and confirm Publish only if it is correct.

Publication creates the version used by new Sessions. It does not rewrite the snapshot already held by an existing Session.

Verify

1. Return to Agents and confirm support-triage is marked published.
2. Start a new Session with that Agent and send the support-report task you used while reviewing the draft.
3. Confirm the response follows the requested shape and that any protected tool action stops for approval.
4. If the behavior is wrong, edit the draft, check it again, and publish a new version. Keep the current Session as the record of what the earlier version did.

Troubleshooting

If the table does not resolve the problem, record the Workspace, Agent ID, Session ID, validation field path, error code, and correlation ID before contacting support. Do not include API keys or other credential material.

Next steps

• Refine Agent behavior when you need exact control over the saved configuration.
• Manage a Session to continue, interrupt, archive, or inspect the Agent’s work.
• For automation, use the Session request shape in Awaken Agents Get Started with agent __admin_assistant, then keep the same review and publication boundary.