How to Achieve Operational Automation with Multi-Agent Systems

Defining agents

from langchain.agents import initialize_agent, AgentType
from langchain.tools import Tool
from langchain.llms import OpenAI

# Data extraction agent
def extract_data(query):
    # logic for querying the SKU database
    pass

extract_tool = Tool(name="extract_sku", func=extract_data)

# Approval agent
approval_llm = OpenAI(temperature=0.1)
approval_agent = initialize_agent(
    [extract_tool],
    approval_llm,
    agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT,
    verbose=False
)

Orchestrating agents

from autogen import AssistantAgent, UserProxyAgent

# Supervisor Agent
supervisor = AssistantAgent(
    name="Supervisor",
    system_message="Coordinate procurement flows, approve high-value orders, and delegate routine tasks.",
    llm_config={"config_list": [{"model": "gpt-4o", "api_key": OPENAI_API_KEY}]}
)

# Routine Agent (low-value orders)
routine_agent = AssistantAgent(
    name="RoutineAgent",
    system_message="Automatically execute routine orders under €500.",
    llm_config={"config_list": [{"model": "gpt-4o", "api_key": OPENAI_API_KEY}]}
)

# Workflow: The user sends a request
user_proxy = UserProxyAgent(name="User", human_input_mode="NEVER")
user_proxy.initiate_chat(supervisor, message="Create an order for 10 units of SKU XYZ.")

Practical example: an automated procurement workflow

1. Stock level detection â€" Agent that queries the ERP database
2. Price analysis â€" Agent that retrieves real-time exchange rates
3. Procurement decision â€" Supervisor agent that evaluates threshold, price, and budget
4. Order generation â€" Execution agent that creates a PDF and sends it to the supplier

Tools and technology stack in 2026

• Orchestration frameworks: LangChain, AutoGen, TemporalAI
• Tool registries: Hugging Face Tool Hub, OpenAI Function Calling
• Monitoring: LangSmith, Arize AI, Grafana for real-time metrics
• Cloud platforms: AWS Bedrock, Azure Cognitive Services, Google Vertex AI

Best practices and security considerations

• Permission limiting: Each agent should have the minimum necessary privileges (e.g., read-only vs. write)
• Traceability: Record each agent decision in an immutable audit log
• A/B testing: Use LangChain's TestableAgents to iterate quickly on workflows
• Security hardening: Implement multi-factor authentication for high-impact agents and use cloud-managed API keys

Conclusion