Build Your First TypeScript AI Agent with Google ADK

Introduction to TypeScript ADK

• 🚀 The TypeScript ADK is an open-source, modular framework designed to simplify the development and deployment of AI agents for TypeScript and JavaScript developers.
• 💡 It addresses the complexity of building agents from scratch, managing state, memory, tool calls, and control flow.

Core Concepts and Agent Types

• 🧩 The fundamental concept is the agent, with three main categories: LLM-based agents (for reasoning and tool calls), workflow agents (for deterministic logic), and custom agents (for custom logic).
• 🛠️ These agent types can be composed together for more advanced scenarios.

Running Your First Agent

• 💻 To get started, include the ADK dependency in your package.json using MPM or YARN.
• 🧪 The ADK includes a dev UI for prototyping and debugging agents with a visual interface, accessible by adding the @google/agent-framework-dev-tools dependency.
• 💬 Programmatically running an agent involves a runner (coordinator) and a session (conversation state), often within a loop to handle user input and agent responses.
• ⚡ Agents can be executed directly from TypeScript source files using ts-node for quick iteration.

Integrating Tools with Agents

• 🌐 Agents can interact with the external world through tools, as LLMs cannot natively browse the internet or query databases.
• 📞 Tools leverage LLM function calling capabilities, allowing agents to determine when an external action is needed, select a registered tool, and pass arguments.
• 🧩 The ADK abstracts the execution loop for tools, providing a clean interface for defining them.
• 🔧 Typescript ADK supports function tools (wrapping your code), built-in tools (like Google Search), and third-party tools.

Advanced Agent Architectures

• 🤝 Multi-agent systems allow multiple agents to collaborate, supported by protocols like MCP and A2A.
• 🔗 A pattern ADK enables is using one agent as a tool for another, facilitating hierarchical or collaborative systems.
• 🔄 Workflow agents (sequential, parallel, loop) are essential for orchestrating complex multi-agent flows, making execution more deterministic.

Context and State Management

• 🧠 Conversational context is managed through sessions, state, and memory.
• 🆔 Sessions manage multiple conversations, storing history, shared state, and interaction events.
• 📝 State acts as a temporary scratchpad for conversation-specific data, requiring serializable values.
• 🎯 Initial state can be specified when creating a session, and agent instructions can access this state using templating syntax for greater generalization.