开发者复盘 Opus 审 400 个 bug 修复：约 1/3 原样采纳，全信就完了
Vjeux · x · 2026-11-13
Mediabunny author vanilagy shares a postmortem on having Opus 5.5 handle nearly 400 fixes. Key takeaways:
- Only about a third of fixes went in unchanged, usually the smaller ones; many left ugly/bloated/repeated/inefficient code, made wrong architectural assumptions or design decisions, or fixed bugs at the wrong level.
- Deciding what even counts as a bug and how the system should behave is something only he, as the project's direction-setter, could add — Opus flagged many issues without fixing them because they required his inspection.
- Opus 5.5 is still "wicked smart," and there are areas he'd blindly trust it.
Core lesson: the better the fixes match your judgment, the more dangerous it is to relax your review discipline.
「编程与Agent」频道最新
- 企业「第二大脑」陷阱:三套系统三个AI,三个过时答案 — alex_verem · 2026-11-14
- Modal 端点接入 Nous Hermes Agent,一行命令装插件 — AAAzzam · 2026-11-14
- Cmpnd 开源 dspy-wsp：花 5-10 美元让 GPT-6 改权重，决策模型全部提分 — dbreunig · 2026-11-14
- AmphoraDB 上线:为 Agent 集群打造的无服务器数据库,钱包即登录、按语句付费 — banteg · 2026-11-14
- 前量化研究负责人辞职做 agent：9 天自动谈下 1000 达人 — Shruti_0810 · 2026-11-14
- 90% 代码 AI 生成后如何守住质量:Limestone 的 V.U.E. 三道门 — alex_verem · 2026-11-14