Agent Development Kit (ADK) Documentation

The Agent Development Kit is Google's open-source, code-first framework for building, evaluating and deploying AI agents, and adk.dev is its documentation hub. Apache-2.0 licensed, it now spans five languages — Python, TypeScript/JavaScript, Go, Java and Kotlin — with the Python package installed as google-adk and requiring Python 3.10 or later; ADK Go 2.0 is the current GA release, adding graph workflows and collaborative agents. The docs are deliberately code-first: a minimal agent is a few lines declaring a name, a model and an instruction string, and everything else layers onto that. Sections cover agent types and managed agents; graph-based workflows with routing, fan-out and fan-in, loops, retries and state management for multi-agent orchestration; a tool ecosystem spanning built-in tools, plain functions, OpenAPI specifications, MCP servers, and other agents used as tools; sessions, conversational memory and state, including context compression to keep token usage bounded; callbacks; observability through logging, metrics and tracing; evaluation; and deployment. The local development loop is a genuine strength — adk run gives an interactive CLI, adk web launches a development UI that can serve a whole directory of agents, adk api_server exposes them over HTTP, and adk eval scores them against test cases. Deployment targets are Cloud Run, GKE and Google's managed Agent Runtime. The framework is model-agnostic, supporting Gemini, Claude, OpenAI models, Gemma, Ollama and vLLM through adapters.

Topic

Agentic

Level

Intermediate

Format

Documentation

Cost

Free

Duration

Reference docs; quickstart ~30 minutes, full agent-to-deployment path ~4-6 hours

Provider

Google

Hands-on

Yes — code/exercises

Certificate

None

• ✓Define an agent from a name, a model and an instruction in code
• ✓Orchestrate multi-agent graph workflows with routing, fan-out, loops and retries
• ✓Attach tools from functions, OpenAPI specs, MCP servers or other agents
• ✓Manage sessions, memory and state including context compression for token budgets
• ✓Run and debug agents locally with the adk run and adk web commands
• ✓Score agents against test cases using the adk eval harness
• ✓Deploy to Cloud Run, GKE or Google's managed Agent Runtime
• ✓Swap models between Gemini, Claude, OpenAI, Ollama and vLLM

• •Five language SDKs — Python, TypeScript, Go, Java and Kotlin — which is unusual among agent frameworks
• •Model-agnostic despite the Google origin: Claude, OpenAI, Gemma, Ollama and vLLM are documented
• •adk web provides a real local debugging UI rather than print-statement tracing
• •Evaluation and deployment are documented alongside the SDK instead of left as an exercise
• •Apache-2.0 and open source; the Python repository alone carries roughly 21,100 GitHub stars

This page was written from 2 sources, 1 on domains other than adk.dev.

1. 1.adk.dev — adk.devvendor
2. 2.github.com — adk python