Lean Agentic AI at the Green Software Development Karlsruhe Meetup

I had the pleasure of delivering a session on Lean Agentic AI at the Green Software Development Karlsruhe meetup, hosted by bluehands Softwareentwicklung.

Connecting SCI to agentic systems

The session focused on connecting the principle of Software Carbon Intensity (SCI) to the way we build agentic systems. Carbon impact is deeply connected to energy usage and token cost — but they are not always the same thing. Token prices are largely uniform regardless of where inference happens; what varies is the grid behind it. Whether the data centre producing those tokens runs on renewable energy or burns coal makes all the difference to the carbon outcome. Optimizing for cost gets you a long way toward optimizing for carbon, but truly lean systems account for both.

Start simple, keep measuring

The encouraging part is that the starting measures are simple — quite similar to the discipline developers already apply to reduce their own usage: right-size the model, trim the context, cap the loops. The real challenge is continuity: keep measuring, keep optimizing, and make the improvement loop part of how the system is engineered rather than a one-time exercise.

For those who want to go deeper, the ideas from the session are covered in my book, Lean Agentic AI. And at the Green Software Foundation, we are working on material that will help teams engineer leaner AI workflows — more on that soon.

Thank you to Aydin Mir Mohammadi and bluehands for hosting, and to the Karlsruhe community for the thoughtful questions and discussion. It’s a wonderful example of what a local software development community can be — I look forward to visiting again.