Sol + Drex 1.5 two-layer agent design splits decisions from execution
A proposed loop uses GPT-6.1 Sol for research and execution and Drex 1.5 for single-pass scoring decisions, with vendor-reported Decision Index numbers comparing Drex 1.5 to Jev.
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 ⭣ ⭣ ⭣
This post quotes another post: x.com/naceai/status/2105319241436328256