We found a new way to build software with agents: the expensive model never writes code

Описание и применение

Автор описывает новую архитектуру разработки, где дорогой модельный агент только наблюдает, а дешёвый выполняет код, что снижает расходы и повышает надёжность.

Задача

Traditional coding agents either cost a lot or require constant babysitting, leading to inefficiencies and unreliable builds.

Результат

The new workflow eliminates the need for the expensive model to write code, reduces token usage to subscription costs (~€100/month), and improves build reliability by separating reasoning, orchestration, and execution.

Рабочий процесс

• 1. Strategist (Claude) runs every 5 minutes, monitors orchestrator state, and decides if work is on track.
• 2. Orchestrator (Orchestro Minimal) holds the backlog, picks next item, hands it to Executor, and records actions.
• 3. Executor (Muse or other cheap agent) performs branch → change → tests → PR → merge if verification is green, then verifies live deployment.
• 4. Independent verifier checks released results and records pass/fail.
• 5. If Strategist detects a project problem, it fixes backlog directly; if a framework problem, it reports to a framework session that updates the orchestrator.

Доказательства

In 14 days my sessions used about $2,100 of tokens at API prices. I paid 100$ in subscriptions.

Источник

The expensive model reads state every 5 minutes; the cheap one does the long runs.

Источник

A real one, from 14 October: the Strategist watching my blog's content pipeline noticed that since we switched to the lighter orchestrator, nothing had been verified for a day.

Источник