Voice agents are having a moment, but building one that actually feels responsive is a different problem from building a text agent. In this workshop, you’ll learn how to think about realtime voice agents in practice — where latency comes from, how memory changes the interaction model, and what it takes to make an agent feel fast enough to talk to.

Misha will walk through a live voice agent built with Mastra and LiveKit, including a new integration that connects voice calls directly into Studio. You’ll see how the agent joins a call, responds in realtime, uses working and observational memory, and handles state without blocking the interaction.

This session will also cover some of the practical tradeoffs that show up once you’re in the weeds: when lower-latency models outperform newer ones, how to record memory asynchronously, and how to avoid common footguns when building voice experiences on top of agent primitives.

What to expect A practical walkthrough of how to build and reason about realtime voice agents with Mastra, with specific discussion of latency, memory, Studio integration, and where LiveKit fits into the stack.

---

Hosted by

• Misha, Engineer, Mastra

  • https://x.com/mnightingale28

  • https://www.linkedin.com/in/michaellitchev/

• Alex Booker, Developer Experience, Mastra

  • https://x.com/bookercodes

  • https://mastra.ai

Recording and code examples will be available to everyone who registers.