Agents Generalize to New Programming Worlds

LLM‑driven coding agents can read a language spec, form hypotheses, run tests, inspect failures, and produce working software even in invented or ecosystem‑free languages (e.g., TeX macros, Brainfuck, a March‑2026 'MNM' candy‑grid language). These are not parroting training data but performing model‑based scientific inference inside novel computational environments. — If agents can adapt to entirely new computational ecosystems without human scaffolding, that accelerates automation, raises new IP/safety questions, and changes who controls digital production and expertise.