Шестнадцать агентов Claude и один Linux: дорогой эксперимент по замене кодеров

Imagine you decided to hire sixteen interns who don't sleep, don't eat, but demand twenty thousand dollars upfront for their work. That's roughly what a recent experiment looked like, in which a group of agents based on Claude attempted to write a full-fledged C language compiler. The result proved ambitious: the system not only produced mountains of code, but also successfully compiled a Linux kernel. However, as is often the case in the modern AI world, the devil lies in the details and the cloud computing bill.

For a long time, we perceived large language models as advanced assistants capable of writing a function to sort a list or finding a bug in a Python script. But creating a compiler is a task of an entirely different order. It's a test of architectural thinking, understanding of low-level processes, and the ability to hold thousands of interconnected relationships in memory. The developers decided not to rely on a single "smart" model, but instead created an entire structure of sixteen agents that interacted with each other, verified code, and corrected errors in real time.

Context here is more important than the fact of code writing itself. The industry is now actively moving from simple chatbots to multi-agent systems. The idea is that if one model makes a mistake, another should correct it. In this case, Claude had to face the harsh reality of systems programming. Creating a compiler required not only text generation, but endless iterations of testing. This is where the twenty-thousand-dollar sum piled up—tokens burned faster than programmers could pour coffee into their cups.

It's important to understand that the magic of "click a button—get a result" didn't happen. The project required deep human management. People served as architects and chief engineers who literally led a swarm of neural networks by the hand through the maze of C language specifications. This doesn't diminish the achievement, but it removes rose-colored glasses from those expecting full automation of development in the next quarter. AI has become a powerful tool, but it still needs a conductor who understands what symphony they're trying to perform.

Why is this important right now? We've reached a point where the cost of intellectual labor by AI is beginning to compete with the cost of labor by highly qualified humans. Twenty thousand dollars for a compiler is expensive for a pet project, but pocket change for a corporation if such a system allows reducing the development cycle by several months. This is a signal to the entire market: the era of "smart suggestions" is ending, the era of autonomous engineering systems is beginning.

Main point: Multi-agent systems are already capable of the most complex systems programming, but for now require human supervision and enormous budgets. Will this become the standard once token costs fall by another tenfold?