Développeurs dépossédés : visiteurs de notre propre code

Developers Left Behind: Inside Our Own Code Visitors

The author warns that abandoning large language models (LLMs) to regain control over their own code could require a massive cleanup. Currently, up to 60% of their project code is unreadable, and this portion is growing with each orchestration. Quitting LLMs is not being considered, as they serve as versatile, non-expert agents that multiply quickly and are useful for tasks where the developer is also novice, such as organizing, planning, and producing team-level office work.

The author describes two distinct approaches to LLM-assisted development: development assisted by LLMs, where the machine generates code but each increment is reviewed by the developer before integration, and vibecoding, where no verification takes place. The author attempted vibecoding three times, but the results were never efficient or readable.

The only successful outcome was radical changes and rewrites, leading to cleaner code. The author concludes that they should stop vibecoding and focus on development assisted by LLMs, with each increment reviewed and validated.

The author discusses the escalation of LLM usage, including the use of MCPs (Multi-Component Prototypes), skills for encoding procedures, and specialized agents. They also mention dangerously-skip-permissions and its use as an invitation to bypass safety measures. The author then describes a complete orchestration via Orca and git worktrees, with multiple agents working in parallel and their work merging while the developer does other tasks.

The author's new role is that of a part-time arbitrator, verifying orchestration success and responding to LLM questions on infrastructure or team matters.