Jev 当决策层：300 个 agent 的果蝇实验面板
视频演示把 Jev 当决策层而非写手：Choice 选路由、Score 打分、Noul 出 0–1 概率，配 Dispatcher 与状态过滤跑 300 个 agent。
Flies swarmed a black Jev tile as a pale room counted three hundred agents. Pale dashboard titled Alba Field tracks Musca core 015 under a locked Jev badge. Bodies, strands, and kinds tick while a red in flight tag stays on. Jev 015 sits in the center as fruit flies orbit the tile in a tight cloud. Core locked text sits in the corner next to three hundred agents. Room panel reads 300 alive while loft numbers flicker. Opacity climbs to 100 percent as marks pass eight thousand. Docket lists six quiet jobs named Gather, Settle, Trace, Mend, Name, and Bind. Job bar moves from zero of six to six of six as plates and strands fill in. Pulse panel draws six channels for compound eye, antenna, thorax, haltere, abdomen, and tarsus. Lines spike while a pink percent tag climbs and falls. Constellation map redraws hot links such as Eye Hal and Ant Thx. Specimen pane holds one live fly on a plate with yaw numbers changing. Agents in the thread treat Jev as a decision layer, not a writer. Language models still draft text while Jev only picks, scores, or answers yes or no. Choice routes among a closed list. Score ranks on a rubric you set, and Noul returns a yes chance from zero to one. State must carry goal, sources, and gaps or the answer stays vague. Playground test comes first so a bad pick is fixed in evidence, not in model hopping. Dispatcher sits between shared state and workers so Jev never writes a briefing. Code then runs the chosen next step while research writes results back. Filter scores old context and drops the rest instead of asking a model to summarize. Ten thousand forks stay cheap because output tokens stay free. Open the thread if you want the ten step split for your own agent loop. Watch the flies again and ask which calls in your stack never needed a sentence.
这条是引用着另一条写的：x.com/bober_smart/status/2103581364134256863