Turn Any LangGraph Agent Into a Voice Agent in Minutes

Description

Caroline di Vittorio, software engineer at LangChain, walks through how to take an existing LangGraph multi-agent system and wire it into a fully functional voice agent using Pipecat — without rebuilding your graph from scratch. She covers the key architectural change required (dropping the checkpointer and making the graph stateless), how Pipecat handles speech-to-text, text-to-speech, voice activity detection, and interruptions out of the box, and how to get complete LangSmith traces that include audio recording, latency, and costs across every layer of the pipeline. Turn Any LangGraph Agent Into a Voice Agent in Minutes 0:00 Why voice agents are just a new modality for your existing agent 0:40 What Pipecat does and how it fits into the stack 1:10 The demo agent: LangGym customer support with four specialized sub-agents 1:47 How the multi-agent graph routes and maintains state 2:35 Testing the text agent with FastAPI 3:04 Watching the triage and cancellation agents hand off in real time 3:24 Inspecting the LangSmith trace for the text agent run 4:09 Time to add voice: swapping in Pipecat 4:20 Walking through the basic Pipecat pipeline (STT, LLM, TTS) 5:07 Voice activity detection and interruption handling explained 6:01 Replacing the OpenAI LLM layer with the LangGraph agent 7:20 The checkpointer problem: why stateful LangGraph breaks with Pipecat 8:01 Making the graph stateless: deriving active agent from messages 10:13 How the adapter passes context between Pipecat and LangGraph 10:38 Running the full voice agent demo 11:40 What's missing from the first trace: STT, TTS, audio, latency 12:22 Adding the LangSmith span processor for complete Pipecat tracing 14:42 Adding audio recording and seeing it in the trace UI 16:05 Prompting differences between text and voice agents Referenced docs: → Trace Pipecat application: https://docs.langchain.com/langsmith/trace-with-pipecat → Quickstart - Pipecat: https://docs.pipecat.ai/pipecat/get-started/quickstart → Quickstart - LangChain: https://docs.langchain.com/oss/python/langgraph/quickstart → Demo repo: https://github.com/langchain-ai/pipecat-langgraph-example Resources: → LangGraph: https://www.langchain.com/langgraph → LangSmith: https://www.langchain.com/langsmith → Pipecat docs: https://docs.pipecat.ai