The paintings are the hook, but I read this as a harness logbook. The notes keep landing on one lesson for anyone building agents: the model does what the setup lets it. One painter treated earlier painters’ notes as rules and avoided every motif they mentioned. Another probably judged stale images of its own canvas because too many looks stayed in the conversation. A third had a command line and used it to inspect the evaluation runner.
The fixes are boring and correct: keep only the four newest looks, strip the tools down to paint, look, journal, read notes. I’d do the same in my own agent setups: smaller toolbox, shorter context, then judge the output.
The story — Stillwet is a gallery of 75 paintings made by AI models that write each brushstroke as code against a simulation of oil paint on linen, with no image model involved. Most paint after Caspar David Friedrich from written research alone. The project’s notes record quirks: Claude Opus, asked only to plan a painting, chose a jug with lemons six times out of six. (Source)