Opus 5.5 推理 + Jev 决策层的研究 agent
作者搭的研究 agent 能读 X 帖、GitHub 仓库、PDF 论文和长视频，Opus 负责推理综合，Jev 控制选源、调查深度、模型与算力选择、工具失败处理和停止时机。
jev + opus 5.5 is becoming one of my favorite ways to build research agents. i built one that can analyze almost any technical source i give it: → x posts and threads → github repositories → pdf files and papers → lectures and long-form videos Opus: handles the actual reasoning and synthesis. JEV: controls the decisions inside the loop: → which source or tool to use next → how deep the agent needs to investigate → which model + effort level makes sense → what to do when a tool fails → when there is enough evidence to stop researching The result is a research agent that can move across completely different sources without running the most expensive reasoning path for every tiny decision Opus thinks. Jev routes. the harness executes and verifies. This is probably one of my favorite agent architectures i've built so far. i break down how to build the decision layer in the article below:
这条是引用着另一条写的：x.com/0xmorlex/status/2101268094387704311