Building Reliable Speech-to-Speech Voice Agents: Combining Gemini Live + LangChain Deep Agents

Caroline di Vittorio, software engineer at LangChain, demos a voice research agent that sidesteps the usual tradeoff between speech-to-speech models and sandwich architectures. The agent uses Gemini Live for natural, low-latency conversation while offloading long-running research tasks to LangChain's Deep Agents, so you get punchy dialogue and accurate, multi-step research at the same time. 

Chapters:
0:00 The tradeoff every voice agent builder faces
1:00 How sandwich architectures work and where they fall short
1:12 Speech-to-speech models: the benefits and the catch
1:38 Introducing the voice research agent
1:58 Best of both worlds: Gemini Live plus Deep Agents
2:22 Live demo: searching Reddit for cake recipes
3:06 How the two agents are orchestrated under the hood
4:03 Swapping in any Deep Agent, including ones that run for minutes
4:30 Solving the blocking problem with async tool acknowledgement
4:55 Walking through the code
6:45 Reading the trace in LangSmith
7:57 Wrap-up and where to find the code

Extra resources
- Deep Agents: https://www.langchain.com/deep-agents
- LangSmith: https://www.langchain.com/langsmith-platform
- Github: https://github.com/langchain-ai/google-adk-realtime-deepagents-example