Four Dots agents coordinate on 386 GitHub issues, 4.8x faster than Codex
The author orchestrated four Dots agents (Blue, Kiwi, Sunny, Berry) to dedupe, reproduce, fix, and test bugs, reporting 4.8x faster and 4.3x cheaper results than Codex working alone on the same queue.
i can't believe what my 4 agents in OpenAI Dots did with 386 GitHub issues I thought they'd just split the backlog. instead, they kept catching what the others missed meet the team: → Blue found repeat reports and related code → Kiwi reproduced the bugs and asked Blue for missing details → Sunny wrote patches and sent unclear cases back to Kiwi → Berry ran the tests and returned broken patches to Sunny one handoff caught a nasty mistake: Sunny fixed an invitation bug, but Berry found that the patch also gave invited users access to billing Kiwi traced the shared permission rule. Sunny changed the patch. the old code failed the new test; the patch passed it across the backlog: 386 issues → 43 reproduced bugs → 12 patches → 9 passing fixes then I compared the same queue two ways: → Codex investigates every issue → 29h 40m → $96.80 → Dot coordinates Blue, Kiwi, Sunny and Berry → 6h 12m → $22.40 same 9 passing fixes. about 4.8x faster and 4.3x cheaper, with every agent call included nothing merged while I slept. I woke up to nine tested diffs and the evidence behind each one take a look at how it works ↓
This post quotes another post: x.com/solomonrojie/status/2105713134174605327