Skip to content

Agent Node

1. Component Intro

The Agent Component is a high-level execution node that triggers specialized AI agents within your workflow. Unlike a standard LLM node that simply generates text, an Agent is often wrapped in a framework (like LangChain) that allows it to use tools, manage its own state, and follow complex, multi-step reasoning paths to solve a specific query.

Core JSON Structure

[[JSON]]

{
 "name": "Agent_Component",
 "type": "agent",
 "description": "Executes a specialized AI agent within the workflow.",
 "output_type": "string",
 "inputs": {
  "agent_id": "agent-123",
  "framework": "langchain",
  "temperature": 0.7,
  "timeout": 30,
  "query": "{{llm_component.output}}",
  "max_retries": 3
 }
}

2. Where to Use It

  • Complex Problem Solving: Tasks that require more than one "thought" step, such as researching a topic and then writing a summary.

  • Tool Orchestration: When you need an AI to decide which API or Database to call based on a user's question.

  • Legacy Framework Support: Integrating existing agents built in LangChain or other supported frameworks directly into the Workflow Weaver.

  • Autonomous Customer Support: Handling requests that involve checking a subscription status, verifying a policy, and then generating a response.

3. How to Initialize

  1. Add Node: Drag the Agent component from the library and place it on your canvas.

  2. Select Agent & Framework: Enter the agent_id and choose the corresponding framework (e.g., LangChain) the agent was built on.

  3. Set Execution Parameters: * Temperature: Adjust the creativity (typically 0.7 for general queries).

  4. Timeout: Set a time limit in seconds (e.g., 30) to prevent the workflow from hanging.

  5. Max Retries: Define how many times the agent should attempt the task if it fails (e.g., 3).

  6. Map the Query: Map the query field to a previous node using {{componentID.output}}.

  7. Format Consistency: Ensure the input and output are connected via String data. If you are receiving JSON from a previous node, you must use an Adapter Component first.

  8. Connect Ports: Link the input dot to the trigger/data source and the output dot to the next processing step.

Agent Node-

Do's and Don'ts

Do's

  • Use the Adapter Node: Always use an Adapter Component before the Agent if your incoming data is in JSON format, as this node strictly requires a string input.

  • Set Realistic Timeouts: Give the agent enough time (at least 30-60 seconds) for complex reasoning tasks that involve multiple tool calls.

  • Monitor Retries: Set a reasonable max_retries (2 or 3) to handle transient network issues without causing infinite loops.

  • Provide Clear Queries: Ensure the query passed to the agent is descriptive; vague queries lead to poor agent performance and wasted tokens.

Don'ts

  • Ignore Data Formats: Don't pass a raw JSON object directly into the query field; the agent will likely fail to parse it or produce gibberish.

  • Over-complicate the Framework: Don't use a heavy Agent framework for simple text transformation; use a standard LLM Node for better speed and lower cost.

  • Hardcode Agent IDs: Don't forget that if you update your agent in the backend, you must ensure the agent_id in the workflow matches the new version.

  • Leave Disconnected: Don't forget that like all logic nodes, the Agent requires a trigger signal via the input port to start execution.

Tip: String-Only Constraints

The Agent component is strictly String-in, String-out.

Why? Most agent frameworks are designed to handle natural language prompts. If your data is currently in a complex JSON object, use the Adapter Component to "Flatten" or "Stringify" that data before it hits the Agent.

Troubleshooting: Timeout Errors

If your Agent node frequently returns a timeout error:

  1. Increase the Timeout: Try moving from 30 to 60 or 90 seconds.
  2. Check Tool Latency: Ensure the tools/APIs the agent is calling are responding quickly.
  3. Simplify the Task: If the task is too broad, the agent may be getting stuck in a reasoning loop.