Skip to main content

Agent node

One AI agent in the app. Connect it to a Brain (which model it uses), Modules (which tools it can call), and the Runtime node (how it runs) on the canvas.

Identity and Prompt​

The Agent panel's Identity and Prompt sections

  • ID (required), the agent's name inside the app. Renaming it updates every place that refers to it, automatically.

  • Role (default assistant) - open text; only the exact value coordinator changes behavior (required for multi-agent delegation without an explicit entry_agent). Anything else is just a label.

  • Specialty (optional), e.g. research, coding.

  • System prompt (required), the agent's core instructions, written once, not per-conversation. Supports {{app.x}}, {{env.x}}, {{secret.x}}, {{prompt.X}}, {{skill.X}}, {{workspace}}. Cannot see live conversation messages or trigger data - it's fixed before the app ever runs. Left empty, the agent still runs on its role/specialty alone.

    In this example the prompt is stored in a separate file (prompts/main.md, referenced as {{prompt.main}}) rather than typed inline - a Convert to inline button switches between the two. The editor itself has Edit / Split / Preview modes, an Improve button, and a live token count.

  • Plan first (boolean, default off).

  • Instructions file (optional), loads instructions from a bundle file instead of system_prompt.

  • Max tool iterations (optional), caps this agent's tool calls per turn, overriding the Runtime node's Max turns.

Modules​

The Agent panel's Modules section, and the module-add palette below the agent card on the canvas

An agent can only use tools it has been explicitly granted. With nothing granted yet, the panel says so directly: "No modules installed. Use + on the agent or Add above." The + on the agent's own canvas card opens a palette of installable modules:

The module palette: Capabilities, Agent, Filesystem, Bash, HTTP, Web, More...

Picking one adds a module node wired to this agent, granting it that module's tools. There is also a Module grants (advanced, JSON) escape hatch on the panel itself for a manual override ({ "filesystem": ["*"], "bash": ["*"] }) - most people won't need it, prefer wiring a module node from the canvas instead.

Delegation and Pool​

The Delegation and Pool sections, and the Brain summary with its Configure shortcut

  • Delegate to - other agents this one may hand off to. + New sub-agent creates one, installs the agent_spawn module, and wires delegation from this agent - confirmed live: with no other agents in the app, the panel explains exactly this rather than showing an empty list.
  • Capabilities (skills) - bundled skill files this agent loads. No bundle skills yet until at least one exists; Open Skills Studio is the shortcut to add one. Per the panel's own note, agents[].capabilities auto-loads skills/*/SKILL.md into this agent's prompt.
  • Pool - worker-pool settings for running several instances of this agent concurrently: Max workers, a Progress updates toggle, and Auto retry (a retry count, default 0).
  • Brain - a read-only summary of the connected Brain node (openai / mimo-v2.5 in this example) with a Configure shortcut that jumps straight to that node's own panel, rather than duplicating its fields here.