OpenAI Dots 跑 4 个 agent 互查修 bug
作者用 OpenAI Dots 一个 dot 驱动 4 个 agent（找 bug、复现、写补丁、跑测试）过夜修完 386 个 issue，比 Codex 单干快 4.8 倍、便宜 4.3 倍。
my ai snitched on my other ai while i was asleep. and honestly it saved my whole app. the patch it caught would've given every invited user access to billing 4 agents. 386 github issues. $22.40. one snitch. full drama below the cast (openai dots, one dot running 4 agents): › blue. the detective. finds repeat reports and digs up the code behind them › kiwi. asks way too many questions. reproduces every bug and keeps bugging blue for details › sunny. the confident one. writes the patches › berry. the snitch. runs the tests and tells on everyone the episode: sunny fixes an invitation bug. clean diff. very proud berry runs one extra test. the "fix" quietly gives every invited user access to billing berry sends it straight back. no meeting, no slack thread, no "lgtm" kiwi tracks it down to one shared permission rule. sunny rewrites the patch, slightly humbled. old code fails the new test. new patch passes by morning: 386 issues → 43 real bugs → 12 patches → 9 that actually pass the part that hurt: › codex working solo: same 9 fixes. 29h 40m. $96.80 › 4 agents checking each other: 6h 12m. $22.40 4.8x faster. 4.3x cheaper. every agent call counted nothing got merged without me. i woke up to 9 tested diffs and receipts for each one i'm keeping berry forever. which one would you fire?
这条是引用着另一条写的：x.com/yurshevv/status/2106054778815983957