Google ADK 2.0

Google ADK 2.0 for Python came out last week and the graph-based workflows are what caught my attention.

I’m currently refactoring an existing ADK architecture to be faster and more reliable. The prompts had grown over time, and coming back to them after working on other things, they were genuinely exhausting to read through. The new graph-based workflows seem like they could address that specific pain point so I’m looking forward to trying them out.

This is also why I keep recommending ADK over just calling Gemini Generative AI endpoints directly. Being able to test and evaluate agent performance matters a lot to me. Updating prompts based on vibes works fine early on, but when you’re already in production, you need to know whether a change made things better or worse.

Breaking an agent into smaller pieces also forces you to focus on the right tool for each step. Not every step warrants an LLM. Some are better served by a different ML algorithm, and some don’t need AI or ML at all.

Whether you’re using Google ADK, LangGraph, CrewAI, or the OpenAI Agents SDK, I think it’s worth exploring at least one.