Skip to main content

Configuring agents

Click an agent's name in the rail to open its configuration. You will find it under Specialists (built-in) or Custom (yours). Every agent, whichever group it is in, has the same three tabs: Behavior, Tools, and Settings.

Behavior​

  • Description: shown to the orchestrator when it is deciding which agent to delegate a piece of work to. This is required for every agent, since it is the only thing the orchestrator has to go on.
  • Rules / prompt: the markdown body of the agent's spec. This is the agent's prompt, not a description of it.
  • What this agent sees: the full context stack in injection order (prompt, shared Project context, this agent's own guidelines, and the Skills catalog if it can spawn sub-agents). See Workspace context for the full breakdown.

Tools​

A searchable, filterable checklist of every tool available to the project. Only checked tools are actually given to the agent.

  • Filter box: matches on tool name.
  • Category dropdown: multi-select; filters by domain (Agents, Sandbox, Board, Web, MCP / Integrations, and so on). Categories come from the tool's own metadata, not a fixed list.
  • Info icon: hover (or focus) any tool to see a popover with its description and parameters (name, type, required or optional).
  • Dormant tools: a tool tied to an integration you have not connected yet is still selectable, but shows an amber dormant badge and stays inert (skipped at build) until you connect that integration in Integrations.
  • MCP tools: tools from a custom-connected MCP server show in their own grouped section at the bottom, by server. See MCP servers.

Settings​

  • Model: leave on (platform default) to inherit the shared model, or pick a specific one from the list the runtime reports.
  • Enabled: a toggle (specialists and custom agents only). Turning it off blocks new work from being dispatched to it; anything already in progress finishes normally. The root orchestrator has no toggle, and always runs.
  • Danger zone (custom agents only): Delete agent, with a confirm step. This removes the spec from the project's docs repo permanently.

How do I create a custom agent?​

  1. Click New agent at the bottom of the rail.
  2. Enter an Agent id: lowercase letters, digits, underscores, and hyphens only (this becomes the role name other agents and orchestrators spawn by).
  3. Enter a Description (required).
  4. Optionally set a Model, pick Tools, and write a Prompt.
  5. Click Create agent.

This creates agents/<id>/profile.md in the project's docs repo.

How do I edit or delete an agent?​

To change any agent's configuration (Specialist or Custom): edit whatever you need across the three tabs, then click Save. A successful save shows Committed — syncing to agents (~1–2 min).

To delete a custom agent:

  1. Open the agent → Settings tab.
  2. Under Danger zone, click Delete agent, then Confirm.

Only Custom agents can be deleted. Built-in Specialists and the root orchestrator cannot be removed; the delete controls simply do not appear for them.

Who can do this: Anyone with project access can view all of the above. Making any change (editing behavior, tools, or settings, creating a custom agent, deleting one) requires agent management access, and only while the project is active and mutable.

On this page