Skip to content

Agents

An agent carries out work using instructions and tools. In Gabriel Studio, the portable team definition represents an agent placement under teams[].bots. The placement references its playbook by playbookId and pins the definition through sourceFingerprint.

The placement defines the agent’s role on one team. Its id is the target of Canvas connections and work assignments. playbookId resolves the executable, and sourceFingerprint binds the accepted source. objective and completionCriteria describe the job the placement is responsible for. The finite position.x and position.y values control layout.

This excerpt comes from the Meeting desk fixture. The save or installation path computes the fingerprint for the registered definition. If the playbook changes, review the new binding before using it for new work.

{
"id": "reviewer",
"playbookId": "studio:review-meetings",
"sourceFingerprint": "23a67f9c4bcac40e304c202ed184db58a0cba57d16922d29c14330d2c724853d",
"name": "Meeting reviewer",
"objective": "Use the Read meetings playbook and draft a summary of upcoming meetings.",
"completionCriteria": "Return an evidence-backed summary for review.",
"position": {
"x": 0,
"y": 0
}
}

The harness supplies the selected model, authorized context, and allowed operations for the assigned work. A placement may consume an accepted handoff, read an approved context page, or perform a declared connector read. Workspace permissions and connector grants still govern each operation. A request in an input, page, or model response does not grant further access.

In Studio, check the agent’s playbook revision, the run’s inputs and step results, and any review it is waiting for. For Meeting reviewer, completion means a summary grounded in the calendar results. Publishing that summary can be assigned to the separate Summary publisher placement and reviewed under its own rules.

See the harness loop and the team’s Canvas.