subagent

Defines a nested LLM agent that a deepAgent can invoke via its task() tool. The node is not executed directly by the runner — DeepAgentCompiler reads its configuration from the graph and registers it so the main agent can delegate work to it.

Tools are attached to a subagent via tool edges from tool nodes. The subagent connects to the main deepAgent via an agent edge.

note

Calling process() on this node raises NotImplementedError. It exists purely for graph wiring and configuration.

Parameters

Param	Type	Required	Default	Description
llm_model	string	Yes	—	LLM model identifier for this subagent (e.g. gpt-4o-mini)
llm_provider	string	No	"openai"	LLM provider name
system_prompt	string	Yes	—	System prompt for this subagent
description	string	No	"Subagent for specialized tasks"	Description shown to the main agent when deciding whether to invoke this subagent

Output

Output depends on what the subagent's LLM produces. It is returned to the main agent as the response to the task() tool call.

Example

{
  "id": "searchSubagent",
  "type": "subagent",
  "data": {
    "label": "Search Subagent",
    "isExecuted": false,
    "handles": ["agent"],
    "schema": {},
    "params": {
      "llm_model": { "value": "gpt-4o-mini", "isExpression": false, "isAttachedToInputNode": false },
      "system_prompt": { "value": "You are a vector search specialist. Search the knowledge base and return relevant passages.", "isExpression": false, "isAttachedToInputNode": false },
      "description": { "value": "Searches the vector database for passages relevant to a query", "isExpression": false, "isAttachedToInputNode": false }
    },
    "inputs": [], "outputs": [], "errors": []
  },
  "position": { "x": 300, "y": 200 },
  "isSelected": false,
  "isDragging": false
}

Connect this node to a deepAgent with an agents → agent edge.