Skip to main content
Sequential patterns in workflow architecture allow you to build multi-agent pipelines where agents execute in a predefined order. Each agent processes data and passes its output to the next agent in the chain.

​
Sequential Execution

In sequential patterns, agents execute one after another in a fixed order. You define explicit edges to control the flow: 
architecture: "workflow"

edges:
  - from: "__start__"
    to: "agent_a"
  - from: "agent_a"
    to: "agent_b"
  - from: "agent_b"
    to: "__end__"
Sequential patterns are ideal when you have a clear, linear workflow and want each agent to build on the previous agent’s output.

​
Core Concepts

​
Special Nodes

  • __start__: Entry point (first agent receives user input)
  • __end__: Exit point (final agent’s output is returned to user)
Every workflow must have edges from __start__ and to __end__.

​
Passing Data Between Agents

Each agent in the sequence receives the previous agent’s output. You can access it in two ways: 1. Direct access (free-form text): The previous agent’s output is automatically available in the agent’s context. 2. Structured output (typed data): Define schemas for agent outputs and reference specific fields in subsequent agents.
Learn more about structured output: Structured Output →

​
Basic Sequential Chain

​
Single Agent Flow

name: "Basic Assistant"
description: "Simple conversational assistant"
architecture: "workflow"
save_messages: true
persistent_state: true

agents:
  - name: "assistant"
    agent_type: "llm_agent"
    description: "A helpful AI assistant"
    system_prompt: "You are a helpful AI assistant."
    
    llm_config:
      temperature: 0.7
      max_tokens: 500
    
    streaming_config:
      show_output_to_user: true

edges:
  - from: "__start__"
    to: "assistant"
  - from: "assistant"
    to: "__end__"

entry_point: "assistant"
Flow: 
START → Assistant → END

​
Sequential Chain with Structured Output

Using structured output for reliable data passing: 
name: "Sentiment Analysis Pipeline"
description: "Analyze → Classify → Respond based on sentiment"
architecture: "workflow"
save_messages: true
persistent_state: true

agents:
  - name: "analyzer"
    agent_type: "llm_agent"
    description: "Analyzes sentiment of user input"
    
    structured_output:
      enabled: true
      schema:
        sentiment:
          type: "str"
          description: "Sentiment category: positive, negative, or neutral"
        confidence:
          type: "float"
          description: "Confidence score between 0.0 and 1.0"
        key_topics:
          type: "list[str]"
          description: "Main topics mentioned in the input"
    
    prompt_config:
      system_prompt: |
        Analyze the sentiment and extract key topics from user input.
    
    streaming_config:
      show_output_to_user: true

  - name: "responder"
    agent_type: "llm_agent"
    description: "Generates response based on sentiment analysis"
    
    prompt_config:
      system_prompt: |
        Sentiment: {{ analyzer.output.sentiment }}
        Confidence: {{ analyzer.output.confidence }}
        Topics: {{ analyzer.output.key_topics }}
        
        Generate an appropriate response based on the sentiment analysis.
        For positive sentiment, be enthusiastic. For negative, be empathetic.
    
    streaming_config:
      show_output_to_user: true

edges:
  - from: "__start__"
    to: "analyzer"
  - from: "analyzer"
    to: "responder"
  - from: "responder"
    to: "__end__"

entry_point: "analyzer"
Flow: 
START → Analyzer (structured output) → Responder (uses analyzer.output.*) → END
The responder agent accesses the analyzer’s structured output using template variables: {{ analyzer.output.sentiment }}, {{ analyzer.output.confidence }}, etc.

​
Research → Analysis Pipeline

Multi-agent workflow where a ReAct agent uses tools to gather data with structured output, then an LLM agent analyzes the findings: 
name: "Research and Analysis Pipeline"
description: "Research with tools → Structured analysis → Insights"
architecture: "workflow"
save_messages: true
persistent_state: true

agents:
  - name: "researcher"
    agent_type: "react_agent"
    description: "Gathers information using web search"
    
    tools:
      - "tavily_search"
    
    tool_calling:
      enabled: true
      max_iterations: 5
    
    structured_output:
      enabled: true
      schema:
        key_findings:
          type: "list[str]"
          description: "Main findings from research"
        sources:
          type: "list[str]"
          description: "URLs of sources used"
        confidence:
          type: "float"
          description: "Confidence in findings (0.0-1.0)"
    
    prompt_config:
      system_prompt: |
        Research the topic using web search. Extract key findings and sources.
    
    streaming_config:
      show_output_to_user: true
      show_reasoning: true

  - name: "analyst"
    agent_type: "llm_agent"
    description: "Analyzes research findings and generates insights"
    
    prompt_config:
      system_prompt: |
        ## Research Findings
        {{ researcher.output.key_findings }}
        
        ## Sources
        {{ researcher.output.sources }}
        
        ## Confidence Level
        {{ researcher.output.confidence }}
        
        Based on the research above, provide:
        1. Key insights and patterns
        2. Recommendations
        3. Potential concerns or limitations
    
    streaming_config:
      show_output_to_user: true

edges:
  - from: "__start__"
    to: "researcher"
  - from: "researcher"
    to: "analyst"
  - from: "analyst"
    to: "__end__"

entry_point: "researcher"
Flow: 
START → Researcher (tools + structured output) → Analyst (uses researcher.output.*) → END
The researcher agent uses tavily_search to gather data and outputs structured findings. The analyst agent then references {{ researcher.output.key_findings }} and other fields to generate insights.

​
When to Use Sequential Patterns

Use sequential patterns when:
  • You have a clear, linear workflow
  • Each step depends on the previous step’s output
  • You want type-safe data passing between agents
  • The execution order is predictable
For dynamic routing, see Supervisor Pattern → or Conditional Edges →

​
Best Practices

Use structured output for data dependencies: When an agent needs specific fields from a previous agent, define a schema to ensure type safety and validation.
Reference outputs explicitly: Use {{ agent_name.output.field }} in prompts to access structured data from previous agents.
Stream intermediate results: Set show_output_to_user: true on agents you want users to see during execution.
Every workflow must have a path from __start__ to __end__. Ensure all agents are reachable.
Keep schemas simple: Start with basic types (str, int, float, bool) before using complex nested structures.

​
Advanced Patterns

For more complex routing logic:

Conditional Edges

Add dynamic routing based on agent output

Supervisor Pattern

Use intelligent task delegation instead of fixed edges

​
Common Use Cases

Pattern: Sequential chainResearch → Draft → Edit → PublishEach agent specializes in one stage of content creation.
Pattern: Sequential chain with toolsCollector (web search) → Analyzer (calculator) → Reporter (synthesis)Each agent uses specific tools for its task.
Pattern: Sequential with structured outputAnalyzer (structured output) → Responder (uses analyzer data)First agent produces typed data, second agent consumes it via template variables.
Pattern: Sequential with conditional edgesWriter → Reviewer → [Approve → Publish | Reject → Writer]Add conditional edges for dynamic routing based on quality checks.

​
Next Steps

Conditional Edges

Add dynamic routing to workflows

Supervisor Pattern

Learn about supervisor-based coordination

Agent Types

Understand different agent types for workflows

YAML Configuration

Complete reference for workflow configuration