Jev 吃掉 agent 62% 算力却不花 token
实测 Jev + Opus 5.5 的 agent：Jev 用 0 token 答完 14300 个分支决策，Opus 只当规划器，上下文零污染。
A MODEL THAT WRITES NO TEXT TAKES 62% OF THE COMPUTE IN AN AGENT BUILT ON OPUS 5.5 The telemetry from a Jev and Claude Opus 5.5 setup looks unusual. Opus gets 38% of the session's compute, Jev gets 62%. Opus spends about 41 thousand tokens and $0.84 per run, while Jev uses 0 tokens and the cost counter shows $0.00. Jev is a decision model of the System One class: it returns an option with a probability, and the code plugs it straight into a condition. Opus 5.5 stays the planner and rewrites the code. The usual way to make AI agents reliable is to send every decision back to the model. Tokens go to choices that have an exact rule, and the model routing lives in a prompt that starts to drift after five edits. Here the loop runs through four stages, and there is no routing prompt in it: > Opus propose - Opus 5.5 proposes a plan or a code change and queries a decision graph where routing and scoring reduce to a single typed call > Jev typecheck - Jev answers every fork in code and returns an option with a probability instead of prose, and untyped requests get blocked (the dashboard shows 413 of them) > Isolated eval - the candidate runs in isolated sandboxes, 32 pods, and the scoring process is inaccessible to its code > Meta distill - the cycle's output is written back into the graph as a new edge, so the next cycle starts with it already in place Every handoff between stages is formalized as a contract: task, inputs, done-condition and confidence. The dashboard shows 2,042 contracts signed, with a handoff latency of 0.42 s. A resolved fork never enters Opus's context. Counter readings: > 14,300 forks answered without a single token in Opus's context > 0.00% context bleed > 0.0% leaks from the sandboxes > +28.4% SOTA gain In the distribution of the last 1000 calls, Opus 5.5 takes 40%, gpt-6.1 22%, grok 5 14% and local Jev 16%. These are the dashboard's own readings and haven't been independently verified. How to split the work between Jev and Opus 5.5 in your own agent is covered in the article below ↓
这条是引用着另一条写的：x.com/model_culture/status/2105354309680435332