Jev 决策文件夹：让 Claude 只写、Jev 拍板、代码执行

Iron Giant @irongiantXBT原推发布于 2026.10.09

在 X 上打开原推

I STOPPED GIVING CLAUDE THE FINAL SAY AFTER BUILDING THIS JEV AGENT FOLDER I used to send every step through Claude: decide, write, act -> now Claude handles the writing. Jev makes the calls, code executes, and every step leaves a receipt inside the folder: • the input > AGENTS.md - when to call Jev and when to skip it > state/build_state.py - goal, workers, done, missing, constraint. grounded in evidence • Jev decides (questions/) > http://route.py - which model tier gets the task > next_worker.py - which worker goes next > http://relevance.py - keep or drop each tool output > http://done.py - has the goal actually been met? > http://risky.py - does this send, pay or delete? • code acts (rules/) > hard_rules.py - stop after ten actions; publishing requires approval > thresholds.yaml - act only when confidence clears the threshold. fraud checks require 0.95 • Claude writes (workers/) > http://research.py - sources and notes > http://writer.py - drafts and briefings > text generation happens only in these workers • the proof > receipts/decisions.jsonl - options offered, chosen ID, re-check and fallback > evals/ - dozens of my own labelled traces, used to tune the thresholds • the guards (hooks/) > pre_tool_use.sh - checks every command before execution > http://stop.sh - verifies "all done" before the agent can stop > 10,000 decisions for $0.42 > 300 ms median latency Claude writes, Jev decides, code acts

这条是引用着另一条写的：x.com/polydao/status/2104783226833186920