Interaction Nodes

Available Interaction Nodes:

LLM Node - Directly interact with Large Language Models
Agent Node - Delegate tasks to pre-configured AI agents

LLM Node

The LLM Node provides direct access to Large Language Models (GPT-4, Claude, etc.) for text generation, analysis, extraction, and decision-making. It’s the most versatile interaction node, supporting chat history, tool calling, and structured output extraction.

If you have existing pipelines with old LLM node format, see the v2.0.0 Migration Guide for updating to the new System/Task structure.

Purpose

Use the LLM Node to:

Generate text based on prompts and context
Analyze content and extract insights
Extract structured data from unstructured text
Have conversations with full chat history support
Call tools via function calling
Make intelligent decisions based on context

Parameters

Parameter	Purpose	Type Options & Examples
System	Provide system-level instructions that set the LLM’s behavior, role, or constraints	`Fixed` - Static system message Example: “You are a helpful assistant.” `F-String` - System message with variables Example: “You are a expert.” `Variable` - System message from state Example: custom_system_prompt
Task	Define the specific task or user request the LLM should process	`Fixed` - Static task Example: “Summarize the text.” `F-String` - Task with embedded variables Example: “Analyze and extract requirements.” `Variable` - Task from state variable Example: task_instruction
Chat History	Provide conversation context from previous interactions	`Fixed` - No history Example: [] `F-String` - Formatted conversation history Example: “Previous conversation: “ `Variable` - Use conversation history Example: messages
Input	Specify which state variables the LLM node reads from	Default states: `input`, `messages` Custom states: Any defined state variables Example: `- input` `- messages` `- user_context`
Output	Define which state variables the LLM’s response should populate	Default: `messages` Custom states: Define specific variables Example: `- extracted_title` `- extracted_description` `- messages`
Toolkits	Bind external tools and MCPs to the LLM for function calling	Toolkits - Service integrations MCPs - Model Context Protocol servers Example: `jira_toolkit:` `- create_issue` `- update_issue` `slack_toolkit:` `- send_message`
Interrupt Before	Pause pipeline execution before this node	Enabled / Disabled Example: `enabled` or `disabled`
Interrupt After	Pause pipeline execution after this node for inspection	Enabled / Disabled Example: `enabled` or `disabled`
Structured Output	Force LLM to return data in structured format matching output variables	Enabled - Response parsed into state variables Disabled - Free-form text to `messages` Example: `true` or `false`

Yaml Configuration

nodes:
  - id: Analyze_feedback
    type: llm
    prompt:
      type: string
      value: ''
    input:
      - input
      - user_context
    output:
      - extracted_title
      - messages
    structured_output: false
    transition: END
    input_mapping:
      system:
        type: fixed
        value: You are a helpful assistant
      task:
        type: fstring
        value: Analyze {user_story} and extract requirements.
      chat_history:
        type: variable
        value: messages
    tool_names:
      JiraAssistant:
        - create_issue
        - update_issue
interrupt_before:
  - Analyze_feedback
state:
  messages:
    type: list
  input:
    type: str
  extracted_title:
    type: str
    value: ''
  user_context:
    type: str
    value: ''

When using structured output with interrupts, include messages in the output variables for meaningful interrupt output.

Select Toolkit/MCP from dropdown
Tool dropdown appears for that toolkit
Select specific tools to make available to the LLM
Repeat for multiple toolkits (each gets its own tool dropdown)

Best Practices

1. Always Include `messages` in Output for Interrupts

When using structured output with interrupts: ✅ Correct:

output: ["extracted_data", "status", "messages"]
structured_output: true

❌ Avoid:

output: ["extracted_data", "status"]  # Missing messages
structured_output: true

2. Use Appropriate Prompt Types

Fixed: For static, unchanging instructions
F-String: When you need to inject specific state variables
Variable: When the entire prompt comes from state

3. Limit Tool Binding

Only bind tools the LLM actually needs: ✅ Good: Select specific relevant tools

toolkits:
  jira_toolkit:
    - create_issue
    - update_issue

❌ Avoid: Binding all tools unnecessarily

toolkits:
  jira_toolkit:
    - [all 50+ tools selected]  # Confuses LLM

4. Structure Your Prompts

Use clear, structured prompts: ✅ Good:

prompt:
  type: "fstring"
  value: |
    ## Task
    Analyze the user story: {user_story}
    
    ## Requirements
    Extract:
    1. Title
    2. Description
    3. Acceptance Criteria
    
    ## Output Format
    Provide structured data for each field.

❌ Avoid: Vague prompts

prompt:
  type: "fixed"
  value: "Do something with the data"

5. Specify Output Variables Clearly

Match output variables to what you’re extracting: ✅ Good:

output: ["jira_project_id", "epic_id", "user_story_title", "messages"]
structured_output: true

6. Use Chat History Wisely

Include messages in input when context matters
Use [] (empty array) for stateless single-turn requests

7. Test with Interrupts

Use interrupts during development to verify LLM outputs:

interrupt_after: true  # Pause after this node to inspect results

8. Handle Tool Calling Errors

When using toolkits, account for potential tool failures in your prompt:

prompt:
  type: "fixed"
  value: |
    Use available tools to complete the task.
    If a tool fails, explain what went wrong and suggest alternatives.

Agent Node

The Agent Node allows you to delegate tasks to pre-configured AI agents that have been added to your pipeline. Instead of configuring LLM behavior from scratch, you leverage existing agents with specialized capabilities, prompts, and toolkits.

Purpose

Use the Agent Node to:

Delegate complex tasks to specialized agents
Reuse existing agents across multiple pipelines
Maintain consistency with pre-configured agent behavior
Simplify workflows by avoiding duplicate LLM configuration
Leverage agent-specific toolkits and integrations

Parameters

Parameter	Purpose	Type Options & Examples
Agent	Select which pre-configured agent to execute	Only agents added to the pipeline appear in dropdown How to Add: 1. Go to Pipeline Configuration > Toolkits 2. Select agents 3. Added agents become available Example: `jira_assistant_agent`
Input	Specify which state variables the agent reads from	Default states: `input`, `messages` Custom states: Any defined state variables Example: `- project_id` `- input`
Output	Define which state variables the agent’s response should populate	Default: `messages` Custom states: Specific variables Example: `- jira_ticket_id` `- ticket_url` `- messages`
Task (Input Mapping)	Map the specific task instruction for the agent	`Fixed` - Static task Example: “Create a Jira ticket for this issue.” `F-String` - Task with variables Example: “Create Jira ticket in : “ `Variable` - Task from state Example: task_instruction
Chat History (Input Mapping)	Map conversation context to provide to the agent	`Fixed` - No history Example: [] `F-String` - Formatted history Example: “Previous context: “ `Variable` - Use conversation history Example: messages
Custom Variables (Input Mapping)	Map agent-specific custom variables if defined	`Fixed` - Static value `F-String` - Value with variables `Variable` - Value from state Example (if agent has `jira_project`): `jira_project: PROJ-123`
Interrupt Before	Pause pipeline execution before this node	Enabled / Disabled Example: `enabled` or `disabled`
Interrupt After	Pause pipeline execution after this node for inspection	Enabled / Disabled Example: `enabled` or `disabled`

Yaml Configuration

nodes:
  - id: Agent 1
    type: agent
    input:
      - input
      - project_id
    output:
      - jira_ticket_id
      - ticket_url
      - messages
    transition: END
    input_mapping:
      task:
        type: fstring
        value: Create Jira ticket in {project_id}
      chat_history:
        type: fixed
        value: []
      project_id:
        type: variable
        value: project_id
    tool: Jiraepam
interrupt_before:
  - Agent 1
state:
  messages:
    type: list
  input:
    type: str
  project_id:
    type: str
    value: ''
  jira_ticket_id:
    type: str
    value: ''
  ticket_url:
    type: str
    value: ''

The Input Mapping section appears after you select an agent. Every agent includes TASK and CHAT_HISTORY mappings. If the agent has custom variables, they also appear as mapping options.

Before using Agent Node, ensure agents are added in Pipeline Configuration > Toolkits section. Only added agents appear in the dropdown.

Best Practices

1. Add Agents to Pipeline First

Ensure the agent is added in Pipeline Configuration > Toolkits section before using Agent Node.

2. Map Task Clearly

Provide clear, specific task instructions: ✅ Good:

task:
  type: "fstring"
  value: "Create Jira ticket in project {project_id} with title '{title}', description '{description}', and priority {priority}."

❌ Avoid:

task:
  type: "variable"
  value: "input"  # Too vague

3. Use Chat History Appropriately

With History: Use when agent needs conversation context
Without History: Use [] for independent, stateless tasks

4. Map Custom Variables Correctly

If agent has custom variables, map them to pipeline state: ✅ Good:

input_mapping:
  jira_project:
    type: "variable"
    value: "project_id"
  sprint_number:
    type: "variable"
    value: "current_sprint"

5. Include `messages` in Output

For debugging and continuity, include messages:

output: ["agent_result", "messages"]

6. Use Interrupts for Testing

Test agent behavior with interrupts:

interrupt_after: true  # Review agent output

7. Reuse Agents Across Pipelines

Create specialized agents once, reuse in multiple pipelines for consistency.

8. Handle Agent Failures

Consider error handling in subsequent nodes:

- id: "check_agent_result"
  type: "router"
  condition: "agent_result is not None"
  routes: ["success_path", "error_path"]