Documentation

Agents

The Agents page (admin console, Global agents) manages the delegatable sub-agent types that are available to every user. These are named personas the main chat agent can hand sub-tasks to: for example a research agent, a coding agent, or a reviewer. The main agent decides when to delegate based on each agent’s When to use text.

Agents come in two groups: built-in agents (defined in code) and custom agents (authored here by an admin). Both are shown as cards; a disabled card is dimmed.

Built-in agents

Built-in agents have hard-coded behavior. Their persona, capabilities, and surface gating are defined in code, so on this page only the on/off toggle and a Default model override are editable. Each card shows the name, a Built-in badge, the description, any required-capability chips, the bound surface (if any), a View prompt expander that reveals the code-defined system prompt, and a Default model picker.

The shipped built-in is Code Reviewer (code-reviewer, gated on the Coding capability). The generic delegation presets Explore and Generic (slugs explore / delegate) also appear as built-in rows: disabling one removes the corresponding explore / generic-delegate tool from runs.

!!! note “Built-ins can’t be deleted” There is no delete action on a built-in card. To remove one from delegation, disable it with the toggle. The only edits a built-in (or integration-contributed) agent accepts are enabled and model; any other field change is rejected.

Custom agents

Custom agents are fully editable. Use Add agent (top right, or the button in the empty state) to create one, and the edit/delete icons on a card to manage it. Each card shows the name, the auto-generated slug badge, the description, and capability/surface chips plus a model chip when a default model is pinned. The toggle enables or disables the agent.

Fields

Field Description
Name Display name. The slug is derived from it automatically; names must be unique among global agents.
When to use (description) Tells the main agent when it should delegate to this agent.
Persona / instructions (instructions) The agent’s system prompt.
Only in surface (requires_surface) Binds the agent to a single surface (chosen from the mode surfaces). Leave as Any surface to allow delegation anywhere.
Required capabilities (required_tools) Multi-select of Web, Devices, Documents, Memory, Coding, Workspace. The agent is only delegatable when all selected capabilities are available in the run.
Default model (model) Optional default model for the sub-agent. Inherit (chat / Auto) (empty) keeps today’s behavior: the worker uses the parent run’s model. Auto (re-resolve) always picks Auto for the worker, or pick a concrete provider:model to pin it.
Enabled Whether the agent is available for delegation.

!!! note “How the default model is applied” The default model is only applied when the chat is on Auto; an explicit chat model always wins, and a per-task model request still overrides it. A pinned model is not a bypass: it still passes the chat’s data-classification and trust-tier gate, so an uncleared pin is rerouted to a governance-compatible model rather than leaking. The sub-agent is gated by the model that actually runs, not the parent’s.

Actions

Action Effect
Add agent Opens the create dialog. Save is blocked until Name is set.
Edit (pencil) Opens the same dialog pre-filled; saving updates the agent.
Delete (trash) Asks for confirmation, then removes the custom agent.
Toggle Enables/disables the agent immediately, without opening the dialog.

Changing a name re-derives the slug; a duplicate name is rejected. A failed save surfaces an error notification and leaves the dialog open.

!!! note “Integration-contributed agents” Agents projected from an installed integration also appear in the global catalog, but only their enabled state and Default model can be changed here - every other field is read-only, since a reconfigure of the integration would overwrite edits. Such rows are created on integration setup and removed when the config entry is deleted.