Agent Node

An autonomous AI agent with tool access. V2 only, Tools Agent type only, no streaming.

Written By pvdyck

Last updated About 5 hours ago

Agent Node

The Agent node creates an autonomous AI agent that reasons through multi-step tasks using connected tools, memory, and a language model.

How It Works

The agent receives a prompt, then iteratively reasons and calls tools until it has enough information to produce a final answer. Each iteration is one "step" — the agent decides which tool to call, interprets the result, and decides the next action.

Parameters

ParameterDescription
PromptThe user message or task for the agent to complete. Supports expressions like {{ $json.input }}.
Require Specific Output FormatWhen enabled, connects an Output Parser to enforce structured JSON output.
Max IterationsMaximum reasoning steps before the agent stops (default: 10).
Return Intermediate StepsInclude the agent's reasoning chain in the output (useful for debugging).
System MessageOptional system prompt to set the agent's persona or instructions.
Automatically Passthrough Binary ImagesWhen enabled, passes image data through the workflow without manual handling.

Sub-Node Connections

InputRequiredDescription
AI Language ModelYesThe LLM powering the agent (e.g., OpenRouter, OpenAI, Anthropic).
MemoryNoConversation memory for multi-turn interactions (e.g., Memory Buffer Window).
ToolNoTools the agent can invoke (Calculator, Code Tool, HTTP Request, etc.).
Output ParserNoEnforces structured output format (V2.1+).
Fallback ModelNoBackup LLM if the primary model fails (V2.1+).

Supported Versions

  • V2, V2.1, V2.2 (default V2.2)
  • Tools Agent type only — Conversational Agent and ReAct Agent types are not supported.
  • V2.1+ adds fallback model and output parser support.

Dynamic Parameters with $fromAI()

Use $fromAI() in connected tool node parameters to let the agent dynamically populate values based on its reasoning, rather than using static configuration. This is available for all tool sub-nodes except the Code Tool.

Limitations

  • No streaming — Responses are returned in full after completion.
  • Tools Agent only — Other agent types (Conversational, ReAct, SQL, Plan and Execute) are not available.
  • Agent tools are limited to supported nodes only (no database or TCP-based tools).

Tips

  • Keep system messages concise — overly long instructions can confuse the agent.
  • Use Return Intermediate Steps during development to understand the agent's reasoning.
  • Connect a Memory Buffer Window for multi-turn conversations.
  • Lower Max Iterations (e.g., 5) if you want faster, cheaper responses.

Related