OpenAI Dots 跑 4 个 agent 通宵修 bug 的绩效复盘

yurshev @yurshevv原推发布于 2026.10.03

在 X 上打开原推

i did performance reviews for my 4 ai employees this morning. one wrote a patch that would've given every invited user access to billing. one snitched on it. payroll for the whole night: $22.40 who gets promoted is at the bottom the team (openai dots, one dot running 4 agents). 386 github issues overnight: blue. finds repeat reports and the code behind them review: quiet, fast, never wrong. meets expectations kiwi. reproduces every bug. if details are missing, kiwi goes back to blue and asks review: annoying in standups. right every time. exceeds expectations sunny. writes the patches review: confident. shipped a "fix" for an invite bug that quietly opened billing to every invited user. needs improvement berry. runs the tests and sends broken patches back review: caught sunny's billing leak before anyone saw it. promoted what happened after berry flagged it: kiwi traced it to one shared permission rule, sunny rewrote the patch. old code failed the new test. new patch passed night totals: 386 issues → 43 real bugs → 12 patches → 9 that pass and then there's codex, the solo guy same queue, same 9 fixes. 29h 40m. $96.80 the team did it in 6h 12m for $22.40 4.8x faster. 4.3x cheaper. every agent call counted codex is on a PIP nothing got merged without my sign off. 9 tested diffs, receipts for each one so who gets promoted? i say berry. sunny is already asking for a raise

这条是引用着另一条写的：x.com/gyome1_/status/2105614581963882963