When you go down a YouTube rabbit hole and come back with a banana squad
I was supposed to be doing something productive. Instead I watched Mark Kashef's video on combining Claude Code agent teams with Nano Banana and thought "I could set that up right now." as I seem to have a habit of doing! 15 minutes later, I had a working agent team generating scored image variants for a workplace mental health graphic. The screenshot below shows the critic agent's ranked output. For anyone unfamiliar, the approach comes from a Google research paper called Paper Banana. The core idea is treating image generation like a design agency: a research agent analyses reference images, a prompt architect creates narrative prompts, a generator calls the Gemini API, and a critic agent scores the results on faithfulness, readability, conciseness, and visual appeal. Then it iterates. What impressed me was the critique loop. Instead of generating one image and hoping for the best, the system produces multiple variants and ranks them against defined criteria. That is genuinely useful for someone without design skills who still needs professional-looking materials. It needs fine tuning for my specific use case, but the foundation is solid. I find with these things that if I do not set them up while I have the headspace, they never get done. So down the rabbit hole I went. Full credit to Mark Kashef for the walkthrough: https://www.youtube.com/watch?v=mGfQV4s1MgE Ryan has had Mark on his channel before so I suspect this will not be entirely unfamiliar territory. But if you have not seen agent teams used for image generation with a built-in feedback loop, it is worth 14 minutes of your time.