What we covered
The session is built around one artefact: a Sri Lanka tourist guide agent that answers by voice, calls tools mid-conversation, and can be cut off mid-sentence without losing the thread. Everyone leaves having built it.
- How AI communication moved from request-response to a persistent, interruptible stream.
- The model ecosystem underneath: Gemini's native audio, and what ADK adds on top of it.
- WebSockets versus WebRTC, and the cases where each is the right transport.
- ADK's architecture: the high-level system, then the tool-calling loop that does the real work.
- Writing it: initialisation, the WebSocket endpoint, custom tool definitions, and interruption handling.
- Observability with LangFuse, so a non-deterministic conversation is still debuggable after the fact.
- Deployment and scaling: what changes between one caller and a thousand.
Format
A hands-on code lab with a public source repository, run as part of DevFest. The final section is a set of open challenges to extend the agent past what we built together.
Who it's for
Developers who want a working real-time agent in front of them by the end of the session, including those meeting ADK for the first time.