Building a Data Science Agent with Google's ADK

Evolution of AI Agents

• 💡 Early generative AI focused on LLMs and prompting, evolving to incorporate RAG pipelines for external knowledge.
• 🚀 The next step involved adding function calling and tools to LLMs, enabling them to perform actions.
• 🧠 Current agents represent a sophisticated evolution, featuring LLM reasoning and planning, tool execution, session management, and orchestration.
• 🤝 The trend is moving towards multi-agent systems, requiring orchestration and communication between multiple agents.

Agent Development Kit (ADK) Overview

• 🛠️ The ADK is a client-side framework and SDK for building sophisticated multi-agent solutions.
• 🧩 It simplifies setting up agents, agent tools, and multi-agent systems, handling sessions and orchestration.
• 🔗 Out-of-the-box tools include function calls and retrieval tools (e.g., for files, AI search, RAG), with the ability to create custom tools.

Data Science Agent Architecture

• 🎯 The Database Agent handles SQL execution against datasets, including SQL validation and fixing.
• 📊 The Data Science Agent performs NL2Py workloads, generates plots, and executes Python code using extensions like code interpreters.
• 🤖 The BQML Agent supports advanced data science tasks, including writing BQML SQL statements for model training and inference.
• 👑 The Root Agent acts as the orchestrator, communicating with the user and routing queries to the appropriate agents or tools.

Codebase and Workflow

• 📂 The codebase follows a pattern with agent.py, instructions.py, and tools.py for each agent, facilitating iteration and customization.
• 📝 Instructions.py centralizes prompts and instructions, allowing for easier prompt tuning and optimization.
• ⚙️ Tools.py defines the tools available to an agent, including agent tools and custom functions like load_artifacts.
• 🧠 Callback contexts (before/after agent calls) allow for pre-computation, such as providing BigQuery schema details to the root agent for efficiency.
• 🔗 The Tool Context acts as a short-term memory for tool-related information, accessible by multiple tools within a session.

Demo and Key Features

• 🚀 The ADK includes a built-in developer front-end (invoked via ADK web) for interacting with multi-agent systems without custom UI code.
• 📈 The demo showcases invoking the Database Agent for data retrieval and the Data Science Agent for plot generation.
• 🤖 The BQML Agent is demonstrated for training forecasting models, highlighting its ability to refine plans and handle errors.
• 🏆 The ADK and its agents can be used for advanced tasks like competing in Kaggle competitions, achieving top rankings with minimal prior experience.