Sol + Drex 1.5 双层 agent 架构:决策层与执行层分离
有人提出用 GPT-6.1 Sol 做研究与执行、Drex 1.5 做单次打分决策的循环架构,并给出厂商自报的 Decision Index 对比数字。
Drex 1.5 > Jev on the Decision Index. the f*cking interesting builder play is GPT-6.1 Sol + Drex 1.5 Sol handles the messy work: research, code, planning and recovery Drex gets a precise decision: score the allowed tools, routes or review options in one pass the loop: observe → Sol proposes → code defines the legal choices → Drex scores → the policy gate checks → tools act → Sol verifies the architecture is the interesting part. the general agent can reason across the whole task. the decision layer gets typed inputs, bounded choices and an outcome you can measure Nace's Drex 1.5 launch kit reports 58.28 on Decision Index 0.2.1, against 57.91 for Jev 1.13.0. it also claims 128K context, 93.4% accuracy on its 32K–128K evaluation and sub-second, single-pass decisions. those are company-reported numbers test it for free ⭣ ⭣ ⭣
这条是引用着另一条写的:x.com/naceai/status/2105319241436328256