Voice Agent observability with LangSmith

Caroline di Vittorio, Engineer at LangChain, builds a voice agent with the Google ADK and the Gemini Live model, then sets up tracing in LangSmith to see exactly what the agent is doing under the hood. Gemini Live is Google's native audio model. It's speech-to-speech, taking audio directly as input and producing audio as output without transcribing to text, which keeps latency low and makes the agent's voice sound natural and emotive.

What's covered: building a terminal-based weather assistant with two tools, defining the LangSmith Google ADK plugin, registering it on the runner, recording user and agent audio so the trace captures only what the user actually heard, and reading the full trace in LangSmith, including audio playback, transcripts, tool calls, interruption events, and token-level cost breakdowns.

Get started with the Gemini Live API. Grab an API key in Google AI Studio and start building.

Resources:
Gemini Live API docs → https://goo.gle/4hDlQYv
Google ADK docs → https://goo.gle/3U4sjlq
LangSmith docs → https://goo.gle/3Tt9CI4

What are you building with Gemini Live? Drop it in the comments.

Subscribe to Google for Developers → https://goo.gle/developers

Speaker:
Products Mentioned: Google AI, ADK,