google-adk

What it is and what it does

ADK is a Python framework for building AI agents using a code-first, graph-based workflow model. It provides two core abstractions—Agent (defines an AI's instructions, tools, and behavior) and Workflow (orchestrates agents and tasks in a directed graph)—plus a runtime engine supporting routing, loops, retry logic, state management, human-in-the-loop steps, and nested workflows. The framework is optimized for Google's Gemini models but remains model-agnostic and can integrate with other LLM providers.

The package bundles a rich tool ecosystem (pre-built tools, custom functions, OpenAPI specs, MCP tools), a development UI for testing and debugging, CLI commands for local execution and evaluation, and deployment support for Cloud Run and Vertex AI Agent Engine. It depends on 25 runtime packages including fastapi, uvicorn, aiohttp, pydantic, and google-genai, reflecting its web-service and async-first design. ADK 2.0 introduced breaking changes to the agent API and session schema; sessions are forward-compatible with ADK 1.28+ but not backward-compatible with older 1.x versions.

Use it for:

• Build multi-turn conversational agents with tool delegation and task routing using the Task API and Workflow runtime.
• Compose hierarchical multi-agent systems where specialized agents delegate work to each other in a graph-based flow.
• Deploy agents to production on Cloud Run or Vertex AI with built-in support for containerization and scaling.
• Evaluate and debug agent behavior locally using the CLI and development UI before deployment.
• Integrate external tools and APIs (OpenAPI specs, MCP tools, custom functions) into agent workflows.
• Implement human-in-the-loop approval flows for sensitive agent actions using tool confirmation.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

A Python framework for building, testing, and deploying AI agents with graph-based workflow orchestration, multi-agent coordination, and tool integration—optimized for Gemini but model-agnostic.

Yes. ADK is actively maintained (last commit 2026-08-14, bi-weekly release cadence), has no known vulnerabilities, and offers a mature, code-first framework for building production-grade multi-agent systems. The permissive Apache 2.0 license and low install friction make it accessible. The 25 runtime dependencies are substantial but well-established packages (fastapi, pydantic, google-genai). Install with the recommended version-specific constraint files to manage transitive dependency stability. Suitable for teams building Gemini-based agents or exploring model-agnostic agent orchestration.

Install

google-adk on PyPI

pip install google-adk

uv add google-adk

poetry add google-adk

Installing google-adk

Before you install

Low friction installation as a pure Python wheel. Active maintenance with releases roughly bi-weekly; last commit 2026-08-14 and 21112 repository stars indicate sustained development. Requires Python 3.10 or later. Documentation recommends installing with version-specific constraint files to manage 25 transitive runtime dependencies.

License in practice

Licensed under Apache 2.0 (permissive), allowing commercial and private use with attribution. No restrictions on modification or redistribution.

Quickstart

pip install google-adk

from google.adk import Agent

agent = Agent(
    name="my_agent",
    model="gemini-2.5-flash",
    instruction="You are a helpful assistant."
)

Requires Python 3.10 or later. Google Cloud credentials and API access needed to invoke Gemini models.

Verify before relying

• Whether the 25 runtime dependencies (aiohttp, fastapi, uvicorn, etc.) introduce meaningful security surface or maintenance burden beyond the package itself.
• Performance characteristics and latency profile for agent execution at scale.
• Whether ADK 2.0's breaking changes from 1.x affect existing production deployments materially.

Package facts

Evidence: google_adk-2.7.0-py3-none-any.whl

Tags

More Python Modules packages

Further reading