The loop your agent can't see

Описание и применение

Модель не может заметить собственный цикл из-за контекстного рота, что приводит к бесконечным пересмотрам дизайна. Решение: использовать безсостояние процессы, хранить состояние на диске, ограничивать ввод, структурировать выводы рецензентов и применять детерминированные лимиты цикла.

Задача

AI development orchestrators can get stuck in a design loop due to context rot, where each model turn inherits errors from previous turns, causing endless revisions and wasted spend.

Результат

The new architecture prevented context rot, eliminated the endless design loop, and reduced wasted model spend. Across 22 production runs, no repeated failures or build failures were observed, and the system reliably moved from design to code.

Рабочий процесс

• 1. Start a new design turn with a fresh model invocation and empty context.
• 2. Load state from a workspace directory (design doc, decisions log, artifact stack).
• 3. Prompt the model with a bounded read‑set: only approved files, targeted symbols, and the previous turn’s verdict file.
• 4. Reviewer writes findings as a structured, schema‑checked file.
• 5. Orchestrator merges findings into a verdict file via plain Python, no model call.
• 6. Next turn reads only the verdict file and relevant design slices.
• 7. Apply deterministic loop caps: halt after three consecutive non‑terminal verdicts or a hard cap on design turns, routing to a human if needed.

Доказательства

The loop cap we'd written never fired.

Источник

We replaced the exempted counter with two that nothing can bypass.

Источник

Across 22 runs and 38 code turns: Same set of change‑owned test failures, turn N and turn N−1 – 0 occurrences.

Источник