Jev + Opus 5.5 做决策层，成本与耗时降约 80%
作者把 Jev 插进 agent 工作流处理上下文筛选、任务分流、失败恢复等小决策，Opus 只负责难推理，整体开销降约八成。
Jev + Opus 5.5 is one of those combinations that makes me wonder why more people aren't building AI workflows this way. In my workflow, adding Jev as a decision layer cut cost and execution time by ~80%. It handles the small, bounded decisions that happen constantly inside an agentic system: which project notes are relevant enough to load into context, whether a task can go to a faster worker, which recovery path to take when a tool fails, and which focused checks should run before triggering the full test suite. Opus still handles the hard reasoning. Jev isn't replacing it. The harness prepares a constrained set of options, Jev makes the cheap decision, and the system validates what happens next. I think this is where a lot of AI workflows are still designed backwards. We keep sending every decision to the most capable model available, even when most of those decisions don't require frontier-level intelligence. The better architecture is to spend expensive intelligence only where the problem is actually expensive to think through. I break down how to build that decision layer in the article below.
这条是引用着另一条写的：x.com/wwardenn/status/2102842288619814925