ADK Streaming Patterns

Comprehensive patterns for configuring Google ADK bidi-streaming (bidirectional streaming) to build real-time, multimodal AI agents with voice, video, and streaming tool capabilities.

Core Concepts

ADK bidi-streaming enables low-latency, bidirectional communication between users and AI agents with:

• Real-time interaction: Process and respond while user is still providing input
• Natural interruption: User can interrupt agent mid-response
• Multimodal support: Text, audio, and video inputs/outputs
• Streaming tools: Tools that yield intermediate results over time
• Session persistence: Maintain context across ~10-minute connection timeouts

Quick Start Patterns

1. Basic Bidi-Streaming Setup

from google.adk.agents import Agent
from google.adk.agents.run_config import RunConfig, StreamingMode
from google.genai import types

# Configure for audio streaming
run_config = RunConfig(
    response_modalities=["AUDIO"],
    streaming_mode=StreamingMode.BIDI
)

# Run agent with bidi-streaming
async for event in agent.run_live(request_queue, run_config=run_config):
    if event.server_content:
        # Handle streaming response
        handle_response(event)

2. LiveRequestQueue Pattern

from google.adk.agents import LiveRequestQueue

# Create queue for multimodal inputs
request_queue = LiveRequestQueue()

# Enqueue text
await request_queue.put("What's the weather?")

# Enqueue audio chunks
await request_queue.put(audio_bytes)

# Signal activity boundaries
await request_queue.put(types.LiveClientRealtimeInput(
    media_chunks=[types.LiveClientRealtimeInputMediaChunk(
        data=audio_chunk
    )]
))

Configuration Patterns

Response Modalities

CRITICAL: Only ONE response modality per session. Cannot switch mid-session.

# Audio output (voice agent)
RunConfig(response_modalities=["AUDIO"])

# Text output (chat agent)
RunConfig(response_modalities=["TEXT"])

Session Management

Session Resumption (automatic reconnection):

RunConfig(
    session_resumption=types.SessionResumptionConfig()
)

Context Window Compression (unlimited sessions):

RunConfig(
    context_window_compression=types.ContextWindowCompressionConfig(
        trigger_tokens=100000,
        sliding_window=types.SlidingWindow(target_tokens=80000)
    )
)

Audio Configuration

See templates/audio-config.py for speech and transcription settings.

Platform Selection

Use environment variable (no code changes needed):

# Google AI Studio (Gemini Live API)
export GOOGLE_GENAI_USE_VERTEXAI=FALSE

# Vertex AI (Live API)
export GOOGLE_GENAI_USE_VERTEXAI=TRUE

Streaming Tools Pattern

Define tools as async generators for continuous results:

@streaming_tool
async def monitor_stock(symbol: str):
    """Stream real-time stock price updates."""
    while True:
        price = await fetch_current_price(symbol)
        yield f"Current price: ${price}"
        await asyncio.sleep(1)

See templates/streaming-tool-template.py for complete pattern.

Event Handling

Process events from run_live():

async for event in agent.run_live(request_queue, run_config=run_config):
    # Server content (agent responses)
    if event.server_content:
        if event.server_content.model_turn:
            # Text/audio from model
            process_model_response(event.server_content.model_turn)

        if event.server_content.turn_complete:
            # Agent finished speaking
            handle_turn_complete()

    # Tool calls
    if event.tool_call:
        # ADK executes tools automatically
        log_tool_execution(event.tool_call)

    # Interruptions
    if event.interrupted:
        handle_interruption()

Multi-Agent Streaming

Transfer stateful sessions between agents:

# Agent 1 creates session
session = await agent1.run_live(request_queue, run_config=config)

# Transfer to Agent 2 (seamless handoff)
await agent2.run_live(
    request_queue,
    run_config=config,
    session=session  # Maintains conversation context
)

Workflows

Complete Bidi-Streaming Agent Workflow

1. Configure RunConfig

   • Choose response modality (AUDIO or TEXT)
   • Enable session resumption
   • Configure context window compression (optional)
   • Set audio/speech configs (for audio modality)

2. Create LiveRequestQueue

   • Initialize queue for multimodal inputs
   • Enqueue messages as they arrive
   • Use activity markers for segmentation

3. Implement Event Handling

   • Process server_content for agent responses
   • Handle tool_call events
   • Manage interruption events
   • Track turn_complete signals

4. Define Streaming Tools (optional)

   • Use async generators for continuous output
   • Yield intermediate results over time
   • Support real-time monitoring/analysis

5. Test and Deploy

   • Validate audio/video processing
   • Test interruption handling
   • Verify session resumption
   • Monitor quota usage

Audio/Video Workflow

See examples/audio-video-agent.py for complete multimodal setup including:

• Audio input processing
• Video frame handling
• Speech configuration
• Transcription settings

Templates

All templates use placeholders only (no hardcoded API keys):

• templates/bidi-streaming-config.py: Complete RunConfig patterns
• templates/streaming-tool-template.py: Async generator tool pattern
• templates/audio-config.py: Speech and transcription setup
• templates/video-config.py: Video frame processing
• templates/liverequest-queue.py: Queue management patterns
• templates/event-handler.py: Event processing patterns

Scripts

Utility scripts for validation and setup:

• scripts/validate-streaming-config.py: Validate RunConfig settings
• scripts/test-liverequest-queue.py: Test queue functionality
• scripts/check-modality-support.py: Verify modality compatibility

Examples

Real-world streaming agent implementations:

• examples/voice-agent.py: Complete audio streaming agent
• examples/video-agent.py: Multimodal video processing agent
• examples/streaming-tool-agent.py: Agent with streaming tools
• examples/multi-agent-handoff.py: Session transfer between agents

Best Practices

1. Choose Modality Carefully: Cannot switch response modality mid-session
2. Use Session Resumption: Prevent disconnection issues
3. Enable Context Compression: For extended conversations
4. Implement Streaming Tools: For real-time monitoring/analysis
5. Handle Interruptions: Natural conversation requires interruption support
6. Segment Context: Use activity markers for logical event boundaries
7. Test Platform Switch: Verify behavior on both AI Studio and Vertex AI

Common Patterns

Pattern 1: Voice Agent with Interruption

See examples/voice-agent.py

Pattern 2: Streaming Analysis Tool

See examples/streaming-tool-agent.py

Pattern 3: Multi-Agent Coordination

See examples/multi-agent-handoff.py

References

• ADK Bidi-streaming Docs
• RunConfig Guide (Part 4)
• Audio/Video Guide (Part 5)
• Real-Time Multi-Agent Architecture

Security Compliance

This skill follows strict security rules:

• All code examples use placeholder values only
• No real API keys, passwords, or secrets
• Environment variable references in all code
• .gitignore protection documented