Give your voice AI agent real abilities (tool calling) @googlecloudtech
Give your voice AI agent real abilities (tool calling)  @googlecloudtech
Uploaded September 2026 | Updated November 2026, 23 hours ago
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
Give your voice AI agent real abilities (tool calling)20 minutes to 2 minutes: How Home Depot automated incident responses with OpenTelemetrySkills vs MCP: How AI tools have evolvedA Developer’s Guide to Gemini 3.5 TranscribeScaling agentic coding practices across your teamHow to onboard to Looker using your favorite LLM and skillsJoin Developers Building with GeminiAI-powered observability & remediation with Google Gemini CLI and DynatraceHow to build and deploy powerful data apps with agentsScale JAX models to multi-GPU systemsCan Gemini transcribe a song?What one agent, everywhere you work actually means
Google Cloud Tech |

Give your voice AI agent real abilities (tool calling)

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER