Definition

The Google Agent Development Kit (ADK) is an open-source, code-first framework for building and orchestrating production-ready AI agents. It is optimized for the Google Cloud ecosystem (Gemini, Vertex AI) but remains model-agnostic and deployment-flexible.

The ADK is built on an async-first, event-driven architecture, treating agents as stateful entities within a managed execution loop.

Core Architecture

Agent Types

  1. LLMAgent: Uses a frontier model (e.g., Gemini 2.5 Pro) to handle open-ended reasoning and dynamic tool selection.
  2. Workflow Agents: Provide deterministic execution using pre-defined patterns (Sequential, Parallel, or Loop).
  3. Custom Agents: Classes that extend BaseAgent to integrate legacy logic or specialized reasoning engines.

Multi-Agent Orchestration

  • Hierarchical Delegation: A "Coordinator" agent can manage multiple specialized "Worker" agents.
  • A2A Protocol: Support for the Agent-to-Agent protocol (via gRPC), allowing ADK agents to interact with agents built in other frameworks.
  • Agent-as-Tool: Modular design where one agent can be exposed as a tool to another agent.

Technical Features

Model Context Protocol (MCP)

The ADK natively supports MCP, allowing agents to act as MCP clients for data retrieval or to expose themselves as MCP tool servers.

Observability and Tracing

Built-in OpenTelemetry instrumentation for tracing every reasoning step, tool call, and state transition. This is the foundation for production monitoring and evaluation.

Human-in-the-Loop (HITL)

Native support for Tool Confirmation flows. The system pauses execution to request manual approval before high-risk tool calls are executed.

Implementation Ownership

Component Responsibility Ownership
Agent Behavior Developer Defining system prompts, workflow graphs, and tool schemas.
Event Loop System The ADK Runner manages message flow and state transitions.
State Persistence Developer Configuring the session store (e.g., Firestore or Redis) for multi-turn memory.
Managed Runtime System Google Vertex AI Agent Engine provides the scaled, managed execution environment.

Examples and Tools

  • Languages: Python, TypeScript, Go, Java.
  • Protocol: MCP (Client/Server), gRPC (A2A).
  • Managed Services: Vertex AI Agent Engine, Google Cloud Run, GKE.
  • Observability: OpenTelemetry, Google Cloud Trace.