Opus 5.5 用 187 行 Python 渲染 agent 成功率模拟

Yumzlef @Yumzlef原推发布于 2026.10.04

在 X 上打开原推

dot "my favorite openai product so far" and says it feels noticeably better every day as it learns more of the workflow and style. To see why small daily gains could feel that big, I simulated 2,000 agent workflows a day for 14 days, 20 steps each, with per-step accuracy rising from 90% to 97.7%. With no checks, a task succeeds only if all 20 steps do: 11% on day 1, 65% on day 14. Add retries when a tool error or a critic catches a wrong step, and it goes from 57% to 89%. Runs where a wrong step slipped through fall from 40% to 11%. The accuracies, the critic and the learning curve are made up. This is not how dot works inside. Claude Opus 5.5 wrote it in 187 lines of Python. It renders in 40 seconds.

这条是引用着另一条写的：x.com/yumzlef/status/2095517043634126889