Skip to main content

Create an Agentic Network

Orchestrate multiple agents to work together as a cohesive unit by building a Agent Network.

In this guide, you will build a network where a Supervisor agent manages two worker agents to fetch stock news and send SMS notifications.

💡 Core Concepts​

Before writing the YAML, it is helpful to visualize how these agents interact.

1. The Architecture​

In a agent network, one agent acts as the Supervisor (the brain), and other agents act as Workers (the hands).

  • Supervisor: Receives the user's request, breaks it down, and delegates tasks.
  • Workers: Execute specific tasks (e.g., searching news, sending SMS) and report back to the Supervisor.

2. The "Edges" (Wiring)​

Edges define the communication flow.

  • Incoming Edge: Who can talk to this agent?
  • Outgoing Edge: Who can this agent talk to?
  • Start Signal: The Supervisor must have "Start" in its incoming_edge to initiate the workflow.

⚙️ Configuration Steps​

Follow these steps to build the Stock News Network from scratch.

Initialize the Network Skeleton​

Define the top-level key create_agent_network and establish the identities of your agents.

create_agent_network:
  agent-1:
    agent_name: "Stock_news_searcher"
  agent-2:
    agent_name: "sms_sender"
  agent-3:
    agent_name: "my_supervisor"
    Role: "supervisor" # Crucial: Marks this agent as the orchestrator

Configure the "Brains" (LLM)​

Every agent needs an LLM configuration to reason and make decisions. You can use different models for different agents, but we will use a consistent config here.

    LLM_config:
        params:
          model: "gpt-5-mini"
          max_tokens: 1000
          request_timeout: 600

info

Pro Tip: You might assign a smarter, more expensive model (like gpt-5) to the Supervisor for complex reasoning, and smaller models to workers for simple tasks.

Equip the Workers (Tools)​

Give your worker agents the specific tools they need to do their jobs.

  1. News Searcher: Needs YahooFinanceNewsTool.
  2. SMS Sender: Needs Twilio_sms.
    # Inside agent-1 (Searcher)
    tools:
      tool_assigned:
        - name: "YahooFinanceNewsTool"

    # Inside agent-2 (SMS Sender)
    tools:
      tool_assigned:
        - name: "Twilio_sms"
          config:
            to: "+1234567890"   # Replace with recipient
            from: "+1987654321" # Replace with Twilio number

info

YahooFinanceNewsTool is currently unavailable. Alternatively, you can use Tavily.

Define Responsibilities (Prompts)​

Use the agent_function to tell each agent exactly what it should do.

  • Supervisor: "Your function is to coordinate the agents defined in outgoing edges (Stock_news_searcher, sms_sender) to answer the user request. Stock_news_searcher for searching the latest news and sms_sender for sending the SMS messages."
  • Searcher: "Your function is to search the web or the internet for the latest news about given companies and the stock market, including all relevant news articles and official filings."
  • Sender: "Your function is to send SMS messages to the user with the given message."
tip

Be explicit in the Supervisor's function about how it should use the workers.

Wire the Connections (Edges)​

This is the most critical step. You must define the flow of control.

  1. Supervisor: Needs incoming_edge: ["Start"] to begin in this scenario. It needs outgoing_edge to both workers.
  2. Workers: Need incoming_edge from the Supervisor and outgoing_edge back to the Supervisor.

The user instructs the Supervisor agent, which is why it is connected to the Start node. The Supervisor must interact with the workers to get the job done.

Each worker only needs to take input from the Supervisor and respond back to the Supervisor.

    # Inside agent-3 (Supervisor)
    incoming_edge:
      - "Start"
      - "Stock_news_searcher" # To receive results back
      - "sms_sender"
    outgoing_edge:
      - "Stock_news_searcher" # To delegate tasks
      - "sms_sender"

📚 Practical Recipes​

Recipe: Stock News & SMS Network​

Use Case: A supervisor coordinates a stock research agent to find stock news and a comms agent to text the summary to a user.

create_agent_network:
  agent-1:
    agent_name: "Stock_news_searcher"
    LLM_config:
        params:
          model: "gpt-5-mini"
          max_tokens: 1000
          request_timeout: 600
    tools:
      tool_assigned:
        - name: "YahooFinanceNewsTool"
    agent_function:
      - "Your function is to search the web or the internet for the latest news about given companies and the stock market, including all relevant news articles and official filings."
    incoming_edge:
      - "my_supervisor"
    outgoing_edge:
      - "my_supervisor"

  agent-2:
    agent_name: "sms_sender"
    LLM_config:
        params:
          model: "gpt-5-mini"
          max_tokens: 1000
          request_timeout: 600
    tools:
      tool_assigned:
        - name: "Twilio_sms"
          config:
            to: "<recipient_phone_number>"
            from: "<twilio_phone_number>"
    agent_function:
      - "Your function is to send SMS messages to the user with the given message."
    incoming_edge:
      - "my_supervisor"
    outgoing_edge:
      - "my_supervisor"

  agent-3:
    agent_name: "my_supervisor"
    Role: "supervisor"
    LLM_config:
        params:
          model: "gpt-5-mini"
          max_tokens: 1000
          request_timeout: 600
    agent_function:
      - "Your function is to coordinate the agents defined in outgoing edges (`Stock_news_searcher`, `sms_sender`) to answer the user request."
    incoming_edge:
      - "Start"
      - "Stock_news_searcher"
      - "sms_sender"
    outgoing_edge:
      - "Stock_news_searcher"
      - "sms_sender"

🚑 Troubleshooting​

  • Workflow never starts

    • Check if the Supervisor agent (Role: supervisor) includes "Start" in its incoming_edge list.

  • Supervisor doesn't call workers

    • Verify the outgoing_edge list in the Supervisor matches the exact agent_name of the workers.
    • Ensure the Supervisor's agent_function clearly describes when to use the workers.

  • "Tool not found" error

    • Ensure the tool names in tool_assigned match the exact system names (e.g., Twilio_sms, not TwilioSMS).