google/adk-python

What it solves

It provides a standardized, code-first framework for building, evaluating, and deploying AI agents. It simplifies the process of orchestrating complex agent workflows, managing multi-agent hierarchies, and integrating diverse tools while maintaining software development principles like testability and versioning.

How it works

ADK uses two primary building blocks: Agent classes to define the AI's instructions and behavior, and Workflow classes to orchestrate these agents using a graph-based execution engine. This engine supports advanced logic such as routing, loops, fan-out/fan-in, and human-in-the-loop (HITL) confirmations. It is model-agnostic but optimized for Gemini, allowing developers to define logic in Python and deploy agents via Docker, Cloud Run, or Vertex AI Agent Engine.

Who it’s for

Developers who want to build sophisticated agentic applications with high control and flexibility, particularly those operating within the Google ecosystem or those needing a structured way to to manage multi-agent delegation and complex state management.

Highlights

• Graph-Based Workflow Runtime: Supports deterministic execution flows with routing, retries, and nested workflows.
• Structured Task API: Enables agent-to-agent delegation with multi-turn task modes and controlled outputs.
• Modular Multi-Agent Systems: Allows the composition of specialized agents into scalable hierarchies.
• Rich Tooling: Integrates with OpenAPI specs, MCP tools, and custom functions.
• Built-in Dev Tools: Includes a development UI for testing and debugging, and a CLI for evaluation.
• Flexible Deployment: Supports containerization via Docker and seamless deployment to Google Cloud.