Agent Development Kit Documentation

What ADK is

ADK is Google's open-source agent framework, released under the Apache 2.0 licence. The Python SDK repository had about 21,600 GitHub stars as of September 2026, with frequent releases (v2.9.2 was published on 18 September 2026). The documentation now lives at adk.dev; the old GitHub Pages address redirects there.

Its stated aim is production agents rather than prototypes, so the docs treat evaluation, context management, observability and deployment as part of building, not an afterthought.

The main concepts in the docs

• Agents: LLM agents defined by instructions, a model and tools.
• Workflow agents: sequential, parallel and loop agents for predictable orchestration, plus graph workflows (new in ADK 2.0) that mix deterministic code with model reasoning.
• Multi-agent systems: agents that delegate to each other, with routing.
• Tools: Python or other language functions, OpenAPI tools and MCP tools, with action confirmations for risky operations.
• A2A protocol: communication between agents, including agents built with other frameworks.
• Sessions, memory and context management: keeping conversation state and compressing or filtering context as it grows.
• Callbacks and events: hooks to inspect or change behaviour at each step.
• Evaluation: test cases, user simulation and environment simulation to check agent behaviour before release.
• Deployment: Google Cloud's agent runtime, Cloud Run, GKE or your own containers, with logging, metrics and traces.

ADK is designed around Gemini but works with other models, including Gemma, Claude, OpenAI models and local models through Ollama or vLLM.

A sensible path through the docs

1. Run the quickstart in your language and open the web development UI, which shows each step, tool call and state change.
2. Build a single agent with two or three tools of your own.
3. Read the workflow agents section and restructure the agent as a sequential or parallel workflow where the steps are predictable.
4. Add evaluation test cases before adding more features.
5. Only then look at multi-agent systems, A2A and deployment.

When ADK is a good choice

• You are on Google Cloud or use Gemini and want first-party tooling and deployment.
• You want built-in evaluation and a debugging UI rather than assembling them yourself.
• Your team works in Java, Go or Kotlin as well as Python.

When to consider something else

• For a simple feature with one or two model calls, the model provider's SDK is less to learn.
• If you are committed to another ecosystem, compare the OpenAI Agents SDK, Anthropic's Claude Agent SDK, LangGraph or Microsoft's agent frameworks.
• Framework abstractions can hide what is sent to the model. Use the dev UI and traces to keep that visible.

Common questions

Is ADK only for Gemini? No. It is optimised for Gemini but supports other models through adapters.

Is it free? The framework is open source. You pay for the models and any cloud services you deploy on.

Does it support MCP? Yes, MCP tools are supported, as is the A2A protocol for agent-to-agent communication.

Best for: developers building production agents, especially on Google Cloud, who want workflows, evaluation and deployment in one framework.