Google ADK

Google’s Agent Development Kit (ADK) is an open-source framework for building agents in Python (with Java support too), designed to make agent development feel like ordinary software engineering. Its core idea is composition: you define an Agent with a model, instruction, and tools, then combine agents into larger systems using workflow agents — SequentialAgent, ParallelAgent, LoopAgent — or dynamic LLM-driven routing where a coordinator delegates to sub-agents. It’s the same framework Google uses inside its own agent products.

ADK’s sweet spot is structured multi-agent systems with a clear path to production on Google Cloud. It’s model-agnostic in principle — you can run Claude, GPT, or open models via LiteLLM — but it’s optimized for Gemini and integrates natively with Vertex AI, where Agent Engine gives you a managed runtime for deployment. Compared to LangGraph’s explicit state graphs, ADK leans on hierarchical agent trees and built-in orchestration primitives; compared to CrewAI, it’s more engineering-oriented, with first-class evaluation, session state, artifacts, and callback hooks. The bundled developer UI (adk web) for tracing and debugging agent runs is a genuine differentiator while you iterate.

In practice you’ll pip install google-adk, define tools as typed Python functions with docstrings (ADK generates the schemas), create agents like Agent(model="gemini-2.5-flash", instruction=..., tools=[...]), and nest them via sub_agents. You test locally with adk web or adk run, wire shared data through session state, and deploy the same code to Cloud Run or Agent Engine when it’s ready.