What's the difference between the Claude Agent SDK and Google's Agent Development Kit (ADK)?

The Claude Agent SDK runs only Claude models and packages Claude Code’s own single-agent loop as a library; Google’s ADK is model-agnostic and built around wiring several agents together into named workflow types instead of running one agent loop. The Claude SDK is literally the Claude Code binary underneath: the same built-in tools, permissions, sessions, hooks, and subagents Claude Code already has, wired into your own Python or TypeScript process. ADK reaches Claude through a native, direct API integration; GPT is native too in some ADK languages (Go ships its own OpenAI package) and reached through the LiteLLM connector in others, including Python. Its Sequential, Parallel, and Loop workflow types still ship as the primary pattern for Java alone; starting with ADK 2.0, Python, TypeScript, and Go replaced them with graph-based and dynamic workflows instead.

That difference decides what you reach for first. Picking the Claude Agent SDK is picking Claude Code’s own loop and tools as a starting point; picking ADK is picking a workflow shape and a model second. The OpenAI Agents SDK draws a similar model-versus-structure line against LangGraph, just with a different pair of frameworks on each side.