Google ADK: Architecture and Essential Components

Google Agent Development Kit (ADK) is a framework designed for creating AI-based agent applications. It offers a modular architecture that enables the development of anything from simple agents to complex multi-agent systems capable of coordinating intricate business workflows. The architecture is composed of several layers, each with a specific function:

1. User interfaces: This top layer includes various channels for user or application interaction with agents, such as CLI, web UI, API endpoints, and custom frontends. This separation allows the agent's logic to remain independent of the user experience, enabling the same agent system to function across multiple channels without requiring internal modifications.

2. ADK Runtime & Runner: This layer is responsible for executing and coordinating the agents' operation. It serves as the operational engine of the architecture and connects user interfaces with the agent system and necessary services. Key components include the Event Loop, Orchestration, and Service Management. The Event Loop manages the sequence of actions during an interaction, while Orchestration controls task execution order and parallel processing.

Service Management connects the Runner to essential infrastructure components like Sessions, Memory, State, Artifacts, Tools, Evaluation services, and Observability systems.

3. Agent System: The core of the architecture, the Agent System allows combining different types of agents based on required autonomy, control, and specialization. There are two main types of agents: LLM Agents and Custom Agents. LLM Agents use language models like those from the Gemini family to interpret requests, reason, make decisions, and potentially execute tasks that don't follow a deterministic flow.

Custom Agents, on the other hand, can incorporate specific business logic, custom algorithms, and complex integrations.

4. Foundation components: These components provide context preservation, action coordination, and interaction with external systems. They include Sessions, which maintain user messages, agent responses, tool calls, intermediate results, and events during execution. Sessions facilitate the transfer of information between agents, enabling continuity and enabling agents to perform tasks beyond simple conversation.