ADK graph workflows for streaming / live voice agents

[arushiag0]

I’m using ADK 2.x graph workflows for deterministic control in text-based agents.

Now I am also trying to build some voice/S2S agents and want to introduce structured control for predictable conversations, but graphs don't currently support live streaming.

Is there any planned support for:

• streaming / event-driven graph execution
• Pipecat-style flows pipelines for voice agents

Or is there a recommended alternative architecture for this use case in ADK?

[arpitraval]

For live voice/S2S agents, I would not model the whole conversation as one blocking graph execution.

A production-friendly pattern is to keep the graph deterministic for state transitions, but make the voice layer event-driven around it.

A practical architecture can look like this:

audio stream / transcript event
  -> turn manager
  -> graph state update
  -> intent / slot decision node
  -> tool decision node
  -> tool execution node
  -> response generation node
  -> streaming output / TTS