LangGraph | AgentDuet SDK

This guide shows how to integrate the AgentDuet SDK with LangGraph for durable agent workflows, including checkpointed state, human-in-the-loop interrupts, and policy-gated tools.

What this integration provides

A voice agent that:

Answers inbound calls and bridges 24 kHz PCM to Gemini Live
Authenticates demo orders against a local mock database (no live store APIs)
Speaks shipping status in one or two short sentences
Allows address changes or cancellations only when fulfillment is unfulfilled

How the pieces fit

Layer	Owns
AgentDuet	Connector, inbound calls, PCM
Gemini Live	Speech-to-speech dialogue + tool calls
LangGraph	Call-scoped order state, tools, fulfillment policy
Your app	Prompts, mock orders, tool side effects

PCM stays on the AgentDuet ↔ Gemini bridge. Graph checkpoints store compact business state only (order id, auth, fulfillment) — not audio.

Prerequisites

Python 3.12+
AgentDuet API key and connector UUID from agentduet.com
Gemini API key for Gemini Live

Step 1: Clone the sample

$ git clone https://github.com/AgentDuet/agentduet-samples.git
$ cd agentduet-samples/integrations/langgraph/order-tracking

Layout:

order-tracking/
├── main.py              # SessionManager + Gemini Live bridge
├── graph.py             # LangGraph OrderSession + tools
├── orders.py            # Mock DB helpers
├── prompts.py           # System prompt + greeting
├── data/orders.json     # Seed orders #1001 / #1002
├── requirements.txt
└── .env.example

Step 2: Install dependencies

$ python3.12 -m venv .venv
$ source .venv/bin/activate   # Windows: .venv\Scripts\activate
$ pip install -r requirements.txt

The sample pins agentduet==1.0.0.

Step 3: Configure `.env`

$ cp .env.example .env

$ AGENTDUET_API_KEY=your-connector-api-key
$ AGENTDUET_CONNECTOR_UUID=your-connector-uuid
$ GEMINI_API_KEY=your-gemini-api-key

Step 4: Run locally

$ python main.py

Call your agent. It should answer and walk order status or a change/cancel.

Demo orders

Order	Zip	Fulfillment	Try
#1001	`94107`	`unfulfilled`	Spoken status; change address or cancel
#1002	`10001`	`fulfilled`	Spoken status; modifications declined

Example scripts:

“Order 1001, zip 94107.” → short spoken status.
“Change my address to 88 Folsom Street, San Francisco, California, 94105.” → succeeds on #1001.
“Order 1002, zip 10001. Please cancel.” → status OK; cancel politely declined.

Step 5: How LangGraph and AgentDuet connect

SessionManager answers the call

From main.py:

1 @sm.on_incoming_call
2 async def on_call(noti: IncomingCallNotification) -> None:
3     session = await sm.open_session(new_session_id(), noti.subscriber)
4     call = await session.process_call(noti)
5     await handle_call(call)

LangGraph owns tools and call state

From graph.py, each call builds an OrderSession with checkpointed OrderState (thread id = call id) and LangChain tools:

authenticate_order — mock DB lookup by order id + zip
check_fulfillment_status — policy gate before modifications
change_shipping_address / cancel_order — allowed only when unfulfilled
hang_up — end the call after the caller says goodbye

Gemini Live receives the tool declarations and dispatches through OrderSession.ainvoke_tool.

Voice prompt rules

The system prompt in prompts.py instructs the agent to:

Greet inbound callers and collect a test order id (#1001 / #1002) plus zip
Authenticate via the mock database tool (never invent status)
Speak shipping status in one or two concise sentences
Check fulfillment before modifications; decline politely when fulfilled
Say a short goodbye and call hang_up when the caller is done
Keep spoken replies short

Notes

Keep call control and PCM buffering on the AgentDuet path; put order policy in LangGraph tools.
Map graph thread ids to stable conversation keys (this sample uses call.id).
Restarting the process reloads seed orders from data/orders.json.
This sample uses 24 kHz audio (CallAudioConfig(sample_rate=24000)).