Google ADK Visual Agent Builder Demo

Multi-agent research system with iterative refinement pattern, built using Google ADK Visual Agent Builder's AI-first approach

A complete demonstration of Google ADK's Visual Agent Builder, showcasing how to build complex multi-agent systems through natural language conversation with an AI Assistant. This research agent uses Google Search with a LoopAgent pattern for iterative query refinement.

📝 Blog Post: Building AI Agents Visually with Google ADK Visual Agent Builder

🎯 What This Demo Shows

This repository demonstrates the AI-first development workflow using Google ADK Visual Agent Builder:

Describe your intent to the AI Assistant in natural language
Generate complete agent architectures automatically
Review the generated hierarchy on the visual canvas
Test immediately with real queries
Deploy production-ready YAML configurations

No manual YAML editing. No configuration errors. Just describe what you want, and the AI Assistant creates the architecture.

🏗️ Architecture

The research agent uses a hierarchical multi-agent pattern:

root_research_agent (ROOT)
└── research_loop_agent (LoopAgent, max_iterations=3)
    ├── query_refinement_agent (LlmAgent, gemini-2.5-pro)
    │   └── tool: google_search
    └── search_execution_agent (LlmAgent, gemini-2.5-flash)
        └── tool: google_search

Key Features:

✅ Iterative Refinement: Loop agent progressively improves search queries
✅ Model Specialization: gemini-2.5-pro for reasoning, gemini-2.5-flash for speed
✅ Smart Termination: exit_loop tool allows early exit when satisfied
✅ Real Google Search: Live integration with Google Search API

📦 What's Included

This demo contains the complete YAML configurations generated by the Visual Agent Builder:

agents/
├── root_agent.yaml                  # Root coordinator (LlmAgent, gemini-2.5-flash)
├── research_loop_agent.yaml         # Loop orchestrator (LoopAgent, max_iterations: 3)
├── query_refinement_agent.yaml      # Query enhancement (LlmAgent, gemini-2.5-pro)
└── search_execution_agent.yaml      # Search & synthesis (LlmAgent, gemini-2.5-flash)

All files are production-ready and can be deployed directly.

🚀 Quick Start

Prerequisites

Python 3.10 or higher
Google Cloud Project with Gemini API access
Google ADK v1.18.0+

Installation

Install Google ADK:

pip install google-adk

Set up Google Cloud credentials:

export GOOGLE_CLOUD_PROJECT="your-project-id"
gcloud auth application-default login

Clone this repository:

git clone https://github.com/thomas-chong/google-adk-visual-agent-builder-demo.git
cd google-adk-visual-agent-builder-demo

💻 Usage

Option 1: Run via ADK Web UI (Recommended)

Experience the Visual Agent Builder interface:

# Launch ADK web interface
adk web

# Open in browser
# http://localhost:8000/dev-ui/

Steps:

Click "+" to create new agent or import agents/root_agent.yaml
Explore the visual canvas showing the agent hierarchy
Test in the chat interface with research queries
View execution trace in the Events tab

Option 2: Run via Python

from google.adk.agents import LlmAgent

# Load the research agent
agent = LlmAgent.from_yaml_file("agents/root_agent.yaml")

# Run a research query
response = agent.run(
    "Research the latest developments in quantum computing error correction in 2024."
)
print(response)

🧪 Example Queries

Try these research topics:

Quantum Computing:

Research the latest developments in quantum computing error correction in 2024.

AI/ML:

Analyze recent breakthroughs in large language model training efficiency.

Edge Computing:

Investigate current trends in edge computing for IoT devices.

Renewable Energy:

Research advancements in solar panel efficiency over the past year.

Expected Behavior

The agent will:

Iteration 1: Generate initial search queries, gather foundational information
Iteration 2: Identify gaps, refine queries, gather deeper details
Iteration 3: Synthesize comprehensive findings, determine if complete
Output: Well-structured research summary with key breakthroughs, comparisons, and future directions

Typical execution time: 60-90 seconds (3 iterations with Google Search calls)

⚙️ Configuration

Change Max Iterations

Edit agents/research_loop_agent.yaml:

max_iterations: 5  # Increase from 3 to 5

Use Different Models

Edit any agent YAML:

model: gemini-2.0-flash-exp  # Try newer models

Customize Research Behavior

Modify the instruction field in agent YAML files:

For more technical depth:

instruction: |
  Focus on technical details, implementation specifics, and academic research.
  Prioritize peer-reviewed papers and technical documentation.

For business focus:

instruction: |
  Emphasize market trends, business implications, and ROI analysis.
  Include competitor analysis and market sizing.

Add More Tools

Extend capabilities in agent YAML:

tools:
  - name: google_search
  - name: url_context      # Fetch content from URLs
  - name: FilesRetrieval   # Analyze local files

📊 How It Works

AI Assistant-First Development

This agent was created by describing requirements to the Visual Agent Builder's AI Assistant:

Prompt:

Create a research agent that uses Google Search with an iterative refinement
pattern. The agent should accept research topics, use a Loop agent pattern to
iteratively improve search queries, and have specialized sub-agents for query
refinement (gemini-2.5-pro) and search execution (gemini-2.5-flash).

AI Assistant Response:

Proposed complete 4-agent architecture
Selected appropriate models (pro for reasoning, flash for speed)
Assigned tools (google_search, exit_loop)
Generated detailed instructions for each agent
Created all YAML files automatically

No manual configuration needed!

Iterative Refinement Pattern

graph TD
    A[User Query] --> B[Root Agent]
    B --> C[Loop Agent Starts]
    C --> D[Query Refinement Agent]
    D --> E[Search Execution Agent]
    E --> F{Satisfied?}
    F -->|No| G[Refine Queries]
    G --> D
    F -->|Yes| H[Exit Loop]
    H --> I[Return Results]

🎓 Learning Resources

From This Demo

Blog Post: Building AI Agents Visually with Google ADK
Architecture Patterns: Study the YAML files to understand multi-agent composition
AI Assistant Usage: See how natural language generates complete architectures

Official Documentation

🛠️ Troubleshooting

"No root_agent found" Error

Cause: Missing or incorrect file paths in YAML

Solution:

# Ensure all YAML files are in the same directory
ls -la *.yaml

# Files should show:
# root_agent.yaml
# research_loop_agent.yaml
# query_refinement_agent.yaml
# search_execution_agent.yaml

API Authentication Errors

Cause: Google Cloud credentials not configured

Solution:

# Login with gcloud
gcloud auth application-default login

# Set project
gcloud config set project YOUR_PROJECT_ID

# Verify credentials
gcloud auth list

Slow Execution

Expected behavior: Loop agents with external tools (Google Search) take 60-90 seconds for 3 iterations.

To speed up:

Reduce max_iterations in research_loop_agent.yaml
Use faster model: gemini-2.0-flash-exp
Optimize search queries in agent instructions

Empty or Poor Results

Check:

Google Search API is enabled in your GCP project
Search queries are specific enough (review query_refinement_agent output)
Loop hasn't exited too early (check iteration count in Events tab)

📈 Performance Notes

Execution Metrics (from demo):

Total time: ~75 seconds
Loop iterations: 3 (max)
Google Search calls: 6-8 searches across iterations
Output quality: Comprehensive 3,000+ word research summaries

Cost Considerations:

Gemini API calls: ~$0.01-0.03 per research query
Google Search API: Depends on your GCP quota
Loop iterations significantly impact cost (fewer iterations = lower cost)

🔧 Advanced Customization

Add Memory/Context

Enable conversation memory:

# In root_agent.yaml
tools:
  - name: google_search
  - name: load_memory
  - name: preload_memory

Create Domain-Specific Variants

Medical Research Agent:

Change search queries to focus on PubMed, medical journals
Add medical terminology to instructions
Increase max_iterations for thorough analysis

Market Research Agent:

Add financial data sources
Focus on trends, market sizing, competitor analysis
Include synthesis agent for SWOT analysis

Code Research Agent:

Add GitHub search tools
Focus on repositories, documentation, code examples
Include code quality assessment criteria

🤝 Contributing

Found a bug or have a suggestion?

Open an Issue
Submit a Pull Request with improvements
Share your customized agent configurations!

📄 License

MIT License - See LICENSE file for details.

Feel free to use this as a starting point for your own research agents!

👤 Author

Thomas Chong

AI Engineer passionate about building practical agent systems and exploring the future of AI-assisted development.

🔗 LinkedIn: linkedin.com/in/chongcht
🐦 X/Twitter: @cch_thomas
📝 Blog: Medium

Have questions or want to discuss agent architectures? Reach out on LinkedIn or X!

🌟 Acknowledgments

Google ADK Team for creating the Visual Agent Builder
Gemini Team for the powerful AI models
Google Cloud for the infrastructure and APIs

🏷️ Tags

google-adk ai-agents multi-agent-systems gemini python agent-development visual-programming llm research-agent loop-agent iterative-refinement google-search

⭐ If you find this demo helpful, please star the repository!

📢 Share your agent architectures - I'd love to see what you build with the Visual Agent Builder. Tag me on X/Twitter @cch_thomas!

Built with Google ADK v1.18.0 Visual Agent Builder Last updated: 2025-11-10

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
agents		agents
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
banner.png		banner.png

License

thomas-chong/google-adk-visual-agent-builder-demo

Folders and files

Latest commit

History

Repository files navigation