The Coding Agent That Can't Grade Its Own Homework

Описание и применение

Фреймворк agent‑loop разделяет процесс генерации кода на несколько этапов с файловыми переходами и независимой проверкой, устраняя самоподтверждение и обеспечивая аудит.

Задача

Coding agents often self‑grade, leading to false “All tests pass” messages and noisy, biased outputs.

Результат

Agent‑loop delivers deterministic, verifiable code changes with a clear audit trail, preventing self‑confirmation bias and ensuring tests are actually run.

Рабочий процесс

• 1. Triage: classify task into S, M, or L tier
• 2. Analyze (M/L): generate analysis.md
• 3. Plan (M/L): produce plan.md with runnable check
• 4. Implement (S/M/L): create implementation.md
• 5. Verify: run independent verifier on plan and diff, output PASS/FAIL
• 6. Debug (on FAIL): up to 3 iterations, each with fresh context and re‑verification

Доказательства

The verifier never sees the reasoning

Источник

The debug loop caps at 3 iterations

Источник