google/adk-python

What it solves

It provides a structured, code-first framework for building, evaluating, and deploying AI agents. It addresses the complexity of orchestrating multiple agents and managing deterministic execution flows, moving agent creation from simple prompts to scalable software engineering practices.

How it works

ADK uses two primary building blocks: Agent (which defines the AI's instructions and tools) and Workflow (which orchestrates agents and tasks using a graph-based execution engine). This engine supports complex logic like routing, loops, fan-out/fan-in, and human-in-the-loop interactions. It is model-agnostic but optimized for Gemini, and allows agents to be defined via Python code or configuration files.

Who it’s for

Developers who want to build sophisticated multi-agent systems with high control over execution flow, testability, and deployment flexibility, particularly those integrating with the Google ecosystem.

Highlights

• Graph-based Workflow Runtime: Supports deterministic flows, state management, and nested workflows.
• Structured Task API: Enables agent-to-agent delegation with multi-turn or single-turn modes.
• Modular Multi-Agent Systems: Allows the composition of specialized agents into hierarchies.
• Rich Tool Integration: Supports custom functions, OpenAPI specs, and MCP tools.
• Built-in Dev Tools: Includes a development UI for testing/debugging and a CLI for evaluation and deployment to Docker or Cloud Run.