	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Anthropic Claude API > Claude Multi-Agent Orchestration

Claude Multi-Agent Orchestration

Author: Venkata Sudhakar

Multi-agent orchestration allows ShopMax India to coordinate several specialized Claude agents working together - one agent handles inventory queries, another processes orders, a third manages customer support - each with its own tools and context, all coordinated by an orchestrator agent.

The Anthropic SDK supports multi-agent patterns by passing tool results between agents. The orchestrator receives a user request, decides which specialist agent to invoke, sends that agent the relevant subtask, collects its response, and either returns the final answer or routes to another agent. Each agent is a separate messages.create() call with its own system prompt and tool set.

The example below shows ShopMax India's orchestrator coordinating an inventory agent and an order agent. The orchestrator delegates tasks using tool calls, each specialist returns structured results, and the orchestrator assembles the final customer response.

import anthropic
import json

client = anthropic.Anthropic()

def inventory_agent(product_id):
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=256,
        system="You are ShopMax India inventory agent. Return stock status as JSON.",
        messages=[{"role": "user", "content": "Check stock for product " + product_id}]
    )
    return {"product_id": product_id, "stock": 45, "warehouse": "Mumbai"}

def order_agent(customer_id, product_id):
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=256,
        system="You are ShopMax India order agent. Process orders and return order ID.",
        messages=[{"role": "user", "content": "Create order for customer " + customer_id + " product " + product_id}]
    )
    return {"order_id": "ORD-9182", "status": "confirmed", "eta": "3 days"}

orchestrator_tools = [
    {
        "name": "check_inventory",
        "description": "Check product stock at ShopMax India warehouses",
        "input_schema": {
            "type": "object",
            "properties": {"product_id": {"type": "string"}},
            "required": ["product_id"]
        }
    },
    {
        "name": "place_order",
        "description": "Place a product order for a ShopMax India customer",
        "input_schema": {
            "type": "object",
            "properties": {
                "customer_id": {"type": "string"},
                "product_id": {"type": "string"}
            },
            "required": ["customer_id", "product_id"]
        }
    }
]

def orchestrate(user_message):
    messages = [{"role": "user", "content": user_message}]
    while True:
        response = client.messages.create(
            model="claude-sonnet-4-5",
            max_tokens=512,
            system="You are ShopMax India customer assistant. Use tools to check inventory and place orders.",
            tools=orchestrator_tools,
            messages=messages
        )
        if response.stop_reason == "end_turn":
            for block in response.content:
                if hasattr(block, "text"):
                    return block.text
        tool_results = []
        for block in response.content:
            if block.type == "tool_use":
                if block.name == "check_inventory":
                    result = inventory_agent(block.input["product_id"])
                elif block.name == "place_order":
                    result = order_agent(block.input["customer_id"], block.input["product_id"])
                else:
                    result = {"error": "unknown tool"}
                tool_results.append({
                    "type": "tool_result",
                    "tool_use_id": block.id,
                    "content": json.dumps(result)
                })
        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": tool_results})

result = orchestrate("I am customer CUST-441 from Bangalore. Is Sony 65-inch TV (PROD-TV65) in stock? If yes, place an order.")
print(result)

It gives the following output,

Great news! The Sony 65-inch TV (PROD-TV65) is in stock at our Mumbai warehouse with 45 units available.

I have placed your order successfully:
- Order ID: ORD-9182
- Status: Confirmed
- Estimated delivery: 3 days to Bangalore

Your order is on its way, valued customer CUST-441!

In production, each specialist agent should have its own retry logic and timeout handling. Keep specialist agent system prompts focused and short - a narrow context window improves accuracy. Use model tiers strategically: the orchestrator benefits from Sonnet's reasoning ability while specialists can run on Haiku for cost efficiency. Log every tool call and result for auditability, especially for order placement operations.

Send your comments, suggestions or queries regarding this site to [email protected].