Code for this episode → https://g.dev/cloud/voicedemo1
Gemini Live API docs → https://g.dev/cloud/gemini-live
Agent Development Kit (ADK) docs → https://g.dev/cloud/adk-docs
Tool calling for real-time voice AI agents, not the request-response version you’ve seen a hundred times. The agent acts mid-conversation and the conversation doesn’t break.
Three parts:
– The loop (declare, decide, execute, return, the model decides, your code acts)
– The same ability written two ways (a raw google-genai dict vs a plain Python function with a docstring in Google ADK)
-And the two rules that decide whether a live agent feels alive or broken, keep every tool instant, and put your policy in one before_tool_callback.
There’s a checkpoint before execution. In ADK it’s before_tool_callback, and it sees every tool call before the tool runs. The demo is a quiet-hours rule: the model tries to play something loud, the callback catches it, the playlist never runs, and the agent still responds in character.
Chapters:
0:00 – Demo: an agent that acts mid-conversation
0:27 – What you’ll learn (three parts)
0:55 – Why the model can only produce words
1:13 – A menu of abilities: name, description, when to use it
1:34 – The loop: declare, decide, execute, return
2:17 – Wrong tool called? Fix the description
2:46 – The raw SDK way: a dict and four arrows you wire
3:07 – The ADK way: a function with a docstring
3:41 – Adding a new ability: set_sleep_timer
4:00 – Demo: the new ability, live
4:18 – Who updates the UI: backend events to the browser
4:56 – Rule 1: tools are synchronous
5:54 – Rule 2: before_tool_callback
6:55 – Demo: the guardrail fires mid-conversation
7:18 – Why this matters when the agent gets a browser
7:51 – Recap: the loop, two ways, two rules
8:29 – Next: controlling a real browser by voice
More resources:
Roadmap preview → https://g.dev/cloud/mma-roadmap
Last episode was Google ADK running the live loop. Next episode the agent controls a real browser, it clicks, it types, it waits for pages to load — which is where this checkpoint stops being optional. → https://g.dev/cloud/real-time-voice-agent
Follow Annie online:
LinkedIn → https://goo.gle/annie-linkedin
X → https://goo.gle/annie-x
Watch more The Omni App → https://g.dev/cloud/modern-ai-app
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
#GeminiLiveAPI #VoiceAI #AIAgents #MULTIMODAL
Speaker: Annie Wang
Products Mentioned: Gemini Live API, Agent Development Kit
