Can AI agents write code to analyze tissue images like biologists do?

Tissue imaging labs spend weeks writing custom code to extract spatial cellular features. CodeCytos treats this as a coding-reasoning problem: a language model reads microscopy data and generates Python to answer questions like "which immune cells cluster near tumor cells?" on the fly. Tested on four expert datasets (brain, lung cancer, pancreas, tonsil), it outperforms baseline methods and works better when trained on generic coding examples than domain-specific ones—suggesting the LLM's raw code-generation skill transfers across tissue types.