Type: url
Original: https://habr.com/ru/articles/1043394/
Status: generated
Language: ru
Summary
A Russian-language Habr case study in which the author describes a month of building Nova, a programming language that expresses side effects in function types, with Claude Code agents executing roughly 300 human-reviewed implementation plans. It classifies four recurring failures of autonomous agents (confident hallucinations, defending an earlier design position, missing a non-standard detail in a typical-looking task, and an echo chamber between agents sharing one context) and the discipline used against them: plans as contracts, worktree isolation, an audit pass by a separately prompted critic, several reviewer agents with different angles, prompts versioned as code, zero test regression and threshold-based escalation to a human.