Why Three Dumb LLMs Are Better Than One Smart One 🧠

Giving one LLM a bunch of tools and telling it to "be a full-stack" is like hiring a junior and assigning them DevOps, backend, design, and investor relations all at once. They'll burn out from head to tail.

When designing AI systems, you can use the Supervisor pattern.
Instead of one god-machine, we create narrow specialists.

For example, we have code for review.
▫️ Agent-StyleChecker: receives the ast tree and looks for bad naming. It has one tool and a strict prompt.
▫️ Agent-BugHunter: looks for eval() and vulnerabilities.
▫️ Agent-DocAdvisor: nitpicks about missing type hints.

Each agent is a separate graph with its own specialized prompt and tools. The probability of hallucination is reduced manifold.
Above them sits a Supervisor. A regular graph node that simply routes tasks: first gives the code to the first agent, then to the second, then to the third, collects results, and outputs a final report. Agents don't even communicate directly with each other — only through the Supervisor. This protects the system from infinite loops where two agents start arguing with each other, burning your API budget.

🗓 Today we'll be breaking down this approach in a live stream at Точка Сборки.

From you — understanding of OOP, basic LLM queries, and how agents work (you can watch recordings of past streams!).

The broadcast and recording will be available to participants. Access is arranged through the bot: https://t.me/TScompiler_bot