Google Agent Development Kit (ADK)

The Google Agent Development Kit (ADK) is an open-source, code-first toolkit from Google for building, evaluating and deploying AI agents and multi-agent systems.

Background. Google announced ADK on April 9, 2025. It is Apache 2.0 licensed, with separate repositories for Python, Java, Go and TypeScript, and documentation that also covers Kotlin. It is optimized for Gemini but works with other models through its model interface. The latest Python release on GitHub at the time of writing is 2.10.0, from September 25, 2026.

Core concepts.

• Agents: LlmAgent uses a model to reason and pick tools. Workflow agents (SequentialAgent, ParallelAgent, LoopAgent) control execution deterministically.
• Multi-agent: agents form a hierarchy through sub_agents, and delegate by model-driven transfer or by calling another agent wrapped as an AgentTool.
• Session, state and memory: a Session holds one conversation’s events and State; a Memory service provides recall across sessions.
• Callbacks and planners: callbacks run custom checks at set points, and planners such as PlanReActPlanner add explicit planning.
• Tool confirmation: an experimental feature that pauses a tool for a yes or no, or for structured input, before it runs.

Evaluation. ADK evaluates both the final response and the trajectory, meaning the tool calls the agent made, against expected values, using .test.json test files for unit-style checks and evalset files for longer multi-turn sessions.

A2A. ADK’s documentation marks its A2A support as experimental, with guides for Python, Go and Java. To expose an agent, to_a2a(root_agent) wraps it as an A2A server and generates its Agent Card. To consume one, RemoteA2aAgent takes a remote Agent Card and lets the remote agent act as a sub-agent. The Python integration works with both a2a-sdk 0.3.x and 1.x, and the docs recommend 1.x for new work.

Neighbouring terms. An agent framework is the general category. ADK’s hierarchy of sub-agents is a common shape for a multi-agent system.