Jev 分工架构：LLM 写、Jev 决策、代码执行

Rhoxbt @Rhoxbt原推发布于 2026.10.24

在 X 上打开原推

WHOEVER LEAKED THIS JEV GRID WITH THE SPLIT SHOULD BE TEACHING THIS AT STANFORD I thought good agents were about picking the smartest model and letting it do everything in one turn. Apparently not. This grid's split is much more interesting: → LLM only writes: it drafts prose and nothing else, no routing, no deciding what the next step is, no "let me think about where this goes" → JEV only decides: Choice, Score, Noul questions against the shared state, output tokens literally zero, 256ms median per decision → CODE only acts: calendar invites fire, mail moves to quarantine, clicks run with outcome checks, no model wake-up anywhere in the loop → ONE STATE is shared across all three roles, 15 walls and 600 cards indexed, no cross-role leaks, every role reads the same world → GATE enforces 0.85 between Jev and code, drops below to review, fires above, flags 6 cases where one role tried to do another's job this run so you get: Writers write. Deciders decide. Acts act. States shared. the interesting part isn't picking the smartest model. it's giving every role one job the other two can't do for it. Instead of one model doing writing, deciding and acting in one freeform turn, you split them: LLM writes, Jev decides, code acts, with 10,000 decisions at $0.42, 256ms median, and gate 0.85 holding every handoff. Write with the LLM. Decide with Jev. Act in code.

这条是引用着另一条写的：「长文：如何搭好 Jev 驱动的 agent 技术栈」