Jev 决策 + Claude 写作的 Agent 文件夹结构
作者公开一套目录：Jev 在 questions/ 里选路由、判相关性、判完成与风险，代码在 rules/ 里兜底，Claude 只在 workers/ 里写，1 万次决策花 $0.42。
MY AGENT MAKES 10,000 DECISIONS FOR $0.42. CLAUDE MAKES NONE i gave Claude one job: write Jev picks. code executes. and every single call gets logged this is the folder that made it work: - what goes in > AGENTS.md - the rulebook for when Jev gets asked and when it doesn't > state/build_state.py - the goal, who's working, what's done, what's missing. raw evidence, no summaries - where Jev picks (questions/) > http://route.py - cheap model or strong model? > next_worker.py - who moves next? > http://relevance.py - is this tool output worth keeping? > http://done.py - is the job really finished? > http://risky.py - is this about to send, pay or delete something? - where code takes over (rules/) > hard_rules.py - max ten actions. nothing goes public without my ok > thresholds.yaml - low confidence = no action. fraud checks need 0.95 - where Claude writes (workers/) > http://research.py - finds sources, takes notes > http://writer.py - drafts and briefs > this is the only corner of the folder that produces text - the paper trail > receipts/decisions.jsonl - what Jev was offered, what it picked, the re-check, the fallback > evals/ - my own labelled traces. every threshold is tuned on them - the bouncers (hooks/) > pre_tool_use.sh - nothing runs until it's checked > http://stop.sh - the agent can't say "done" until that's verified 300 ms median per call. every decision on record. Claude writes. Jev decides. code acts.
这条是引用着另一条写的：x.com/thegreatest_sv/status/2102802378378621355