Sobmit Docs

Agent node

Agent node card

The Agent node calls a language model through the AI gateway. It is the core “brain” of most flows.

Handles

HandlePurpose
InMain line — the payload becomes the prompt (unless prompt overrides)
promptInstruction override — replaces the payload-derived prompt
fileFiles — images and PDFs are sent as model file inputs

Configuration

  • Model — the gateway model id (e.g. openai/gpt-5.6-luna).
  • Instructions — a system prompt that always applies.
  • Output — text (free-form) or json-schema (structured).

Structured output

In json-schema mode you describe the response shape. The schema root must be { "type": "object" } with at least one property. Example schema:

{
  "type": "object",
  "properties": {
    "label": { "type": "string" },
    "confidence": { "type": "number", "description": "0 to 1" }
  },
  "required": ["label", "confidence"]
}

Advanced settings

  • Temperature — randomness (empty = provider default).
  • Max tokens — response length cap.
  • Seed — deterministic output.
  • Input from — dot-path to a field of the incoming payload to use as the prompt instead of the whole payload.

Example — classification

Prompt node → Agent (json-schema output with label + confidence) → Router (branches on label).

Notes

  • A model without vision cannot read image inputs — the node warns when you wire an image source into it.
  • PDFs require a model with document input (the gpt-4.1 family works).
  • Structured output and tool calling cancel each other out on some models — Sobmit uses a two-call pattern (research, then extract) when both are needed.

Related