Frontend developer Ankur Sethi spent four weeks in January and February 2026 building Cutlet, a fully functional dynamic programming language, using Anthropic's Claude Code agentic coding tool — without reading a single line of the generated C code himself. Sethi, who has used LLM-assisted programming since GitHub Copilot's original 2021 release, designed a structured set of guardrails to enable autonomous iteration: comprehensive test suites, detailed specification documents, and Docker-based feedback loops that let Claude Code self-correct by running make test until errors disappeared. The resulting language, named after his cat, supports arrays, maps, a novel @ meta-operator for vectorized and reduce operations, prototypal inheritance, mixins, a mark-and-sweep garbage collector, and an interactive REPL — running on both macOS and Linux.
Sethi frames the experiment as a demonstration of "agentic engineering" — a discipline he argues requires human skill in planning, specification design, and feedback environment construction rather than direct code authorship. He identifies four core competencies for effective agentic work and is candid about where the approach breaks down: LLMs excel at well-precedented, algorithmically verifiable problems like interpreter construction, where test suites provide clear correctness signals, but struggle with novel visual and design tasks that lack objective criteria. His own frontend work — complex data visualizations and responsive animations — has resisted LLM assistance precisely because such tasks cannot be reduced to pass/fail tests.
The post drew pointed debate on Hacker News around <a href="/news/2026-03-14-nyt-ai-coding-assistants-end-of-programming-jobs">code ownership and what programming languages are actually for</a>. Commenter andsoitis raised the paradox that programming languages exist as human-computer interfaces — if no human writes or reads the code, the language itself becomes irrelevant as a medium. Others pushed back: commenter bobjordan argued that while LLMs enable codebases far larger than any single human could build unaided, human strategic direction remains indispensable for exploring problem spaces and maintaining coherence across hundreds of micro-plans. The prevailing consensus landed on "tools don't own the house" — authorship stays with the human architect regardless of who swings the hammer.
Sethi's framing holds: building Cutlet required four weeks of deliberate human work, just not the kind that produces C files. The investment was in feedback environment design, specification clarity, and knowing which problems suit autonomous generation — and which don't.