Jev 决策路由层技术笔记：31,000 条边、94.4% 成功率
独立研究者发布 Jev 技术笔记，称其作为多 agent 图上的类型化决策路由层，94.4% 任务成功率、中位延迟 1287ms，微软 agent 团队 6 小时内收藏。
independent researchers just dropped a technical note that microsoft's agent team bookmarked within 6 hours - JEV Engineering, a decision routing layer that routes 31,000+ live edges across multi-agent graphs with 94.4% task success rate and 1,287ms median end-to-end latency that's the architecture that decides WHICH AI agent does WHAT, in real time, with confidence scores, policy gates, and human escalation built in. the routing formula is r* = argmax P(r|s) across a scored candidate set that rebuilds every single step. no static pipelines. no hardcoded chains. pure dynamic selection. JEV is not generative. it's a typed decision layer - Choice (up to 255 options), Score (rank), Noul (no valid route). one thing better than any LLM: route. the graph model G = (V,E) where each node is an agent, model or tool and each edge is a weighted transition with cost, latency, and success probability encoded. the topology is dynamic - new nodes and edges added as tasks progress, low-score routes pruned automatically. it self-heals. it backtracks. it escalates to humans when confidence drops below threshold. 32k context window. 70-500ms typical call latency. 94.4% task success vs 70.8% lexical baseline. the numbers aren't close. who needs this right now - Palantir, Scale AI, any enterprise deploying more than 3 AI agents simultaneously. the TAM for agent orchestration infrastructure is $47B by 2028. JEV is the missing routing layer every serious deployment needs before they can scale past 10 concurrent agents without collapse. built outside of any institution. one team. no VC backing listed. October 2026. the people who find infrastructure papers like this before the press does show up to the meeting already knowing the answer. Bookmark & watch today. The graph doesn't route itself.
这条是引用着另一条写的：x.com/0xdeliriumm/status/2091560754122014781