Building voice agents can come with tradeoffs.
💬 Better convos w/ speech to speech models Vs 🥪 More reliable harnesses w/ sandwich architecture
How to build a voice research agent with both: ✅ Gemini Live: low-latency, natural-sounding convos ✅ Deep Agents: long-running research tasks ✅ LangSmith: full tracing & observability
Users praise the clever workflow for managing subagents when building voice research agents with Gemini Live.
No Digg Deeper questions have been answered for this story yet.
Tutorial on using Gemini live to build a voice agent
Uses deepagents as a tool: offload complex work to this subagent, use Gemini live for the naturalness/latency
Building voice agents can come with tradeoffs.
💬 Better convos w/ speech to speech models Vs 🥪 More reliable harnesses w/ sandwich architecture
How to build a voice research agent with both: ✅ Gemini Live: low-latency, natural-sounding convos ✅ Deep Agents: long-running research tasks ✅ LangSmith: full tracing & observability
Watch the full video on YouTube: https://www.youtube.com/watch?v=aV0ctbnDUCM
@LangChain Native multimodal models are emerging. This architecture might become a temporary solution as models evolve.
@hwchase17 that workflow for managing subagents is actually super clever.
@LangChain The sandwich approach routes audio through ASR → LLM → TTS. Each hop adds roughly 300-500ms. That's where latency originates.
@LangChain In research tasks, a few hundred ms delay is negligible compared to hallucinated results. Reliability matters more than speed.