Gemini Interactions API Samples

This directory contains reference implementations of AI Agents refactored from the Google Agent Development Kit (ADK).

These samples demonstrate the "Gemini 3" / Native Agentic architecture, where the orchestration logic (loops, state management, tool selection) is offloaded from client-side frameworks to the model itself via the Interactions API.

🚀 Setup

1. Install the SDK This requires the new google-genai library (v1beta).

   pip install google-genai

2. Set your API Key You need a Gemini API key.

   export GEMINI_API_KEY="your_api_key_here"

📂 The Agents

1. Customer Service Agent (interactions_customer_service.py)

• Concept: Stateful, transactional bot with many tools.
• Key Feature: Demonstrates Automatic Function Calling. The model autonomously calls tools like check_product_availability and modify_cart to solve user requests.
• Run:

  python interactions_customer_service.py

2. Software Bug Assistant (interactions_software_assistant.py)

• Concept: Integration Agent connecting to external systems (Mocked).
• Key Feature: Demonstrates API Integration. The agent acts as a bridge between a user and systems like a Ticket Database or StackOverflow.
• Run:

  python interactions_software_assistant.py

3. Data Science Router (interactions_data_science.py)

• Concept: Hierarchical / Routing Agent.
• Key Feature: Demonstrates Native Reasoning. The model decides whether to query a SQL database (BigQuery) or perform Python analysis, without hardcoded "Router" logic. Watch the [Root Agent Reasoning] logs.
• Run:

  python interactions_data_science.py

4. Red Teaming / Security (interactions_red_teaming.py)

• Concept: Multi-Agent Orchestration.
• Key Feature: Demonstrates Workflow Planning. A single "Orchestrator" model manages a team of virtual sub-agents (Red Team, Target, Evaluator) by calling them as functions.
• Run:

  python interactions_red_teaming.py

5. RAG / Knowledge Agent (interactions_rag_agent.py)

• Concept: Retrieval Augmented Generation.
• Key Feature: Demonstrates the Production Gap. While the model knows when to search, the infrastructure code (retrieve_rag_documentation) must still be implemented by the developer using frameworks or libraries.
• Run:

  python interactions_rag_agent.py

📝 Key Learnings for "Agentic Reasoning"

• Logic: Moved from Python (while loops) to Model (output.thought).
• State: Moved from Python (messages=[]) to API (interaction_id).
• Code: Significantly reduced boilerplate compared to framework-based implementations.