Mathieu Acher, a computer science professor at INSA Rennes, has been running experiments that suggest recent coding agents are starting to look less like autocomplete systems and more like adaptable problem-solvers: systems that can enter unfamiliar computational worlds, infer the rules, and build substantial working software with only light human guidance.He asked Claude Code (Opus 4.6) and Codex CLI (GPT-5.2) to build chess engines from scratch across 12 languages, including ones where no chess engine has ever existed: LaTeX, Brainfuck, Rocq. No architecture docs, no step-by-step guidance. The results range from ~900 to ~2200 Elo depending on language. The TeX engine alone is ~2,100 lines of pure macro expansion, using \count registers as RAM and \csname tables as ROM (there's nothing like it to copy from).His latest experiment pushes the idea further. He challenged Claude Code with MNM Lang, an esoteric language invented in March 2026 in which source code is represented as colored M&Ms arranged on a grid. With essentially no ecosystem beyond the language spec and opcode table, the agent solved all 26 challenges, culminating in a working Brainfuck interpreter written in MNM Lang.The key distinction, in Acher’s telling, is that these are not bare LLM completions but agentic systems that read specs, form hypotheses, run code, inspect failures, debug, and iterate. His broader claim is that the agent loop gives LLMs a real capacity to adapt to novel computational environments rather than merely reproducing familiar patterns from training data.Sources:1. Coding Agents Build Chess Engines From Scratch in Rust, C++, COBOL, Rocq, LaTeX, Brainfuck, and More https://blog.mathieuacher.com/FromScratchChessEnginesPolyglot/2. TeXCCChess: How Coding Agents Wrote a Chess Engine in Pure TeX https://blog.mathieuacher.com/TeXCCChessEngine/3. Can Coding Agents Program in M&Ms Language? https://blog.mathieuacher.com/CodingAgentsMnMLang/