用 Jev 当裁判的 6 步 agent 生产循环
作者给出「LLM 写、Jev 判、代码决定」的 6 步循环：状态打包、Jev 并行回答多个 typed 问题、按分支起草、验证、三档门控、留可回放回执。
99% of AI agents in production are fundamentally broken one model writes the plan, approves its own tool calls, evaluates its own output, and ships it. zero audit trail. zero safety. the stack that fixes it: llm writes. jev judges. code decides. here is the exact 6-step loop: → state - code builds the smallest evidence packet and strips secrets → ask - jev answers several typed questions at once (Choice, Score, Noul) over the same state → draft - the llm only sees the context for the branch jev picked → verify - parsers, tests and allowlists run first, then jev checks narrow properties of the output → gate - thresholds turn probabilities into three outcomes: auto, review, stop → receipt - state, question version, model, route and outcome get logged and can be replayed what makes it different: • a typed answer comes with a probability, so uncertainty is measurable • one request can carry many questions (TypeSafe's cookbook reports roughly 10x faster than separate calls in a 13-question example) • rollout is staged: offline → shadow → assist → limited → expand the breakthrough insight: a confident answer is not permission. when jev lands in the middle band, the harness asks for more evidence or a human instead of letting the model guess. the engineer who ships agents with a kill switch and a receipt for every decision is the one who gets trusted with production. save this architecture for your next agent build.
这条是引用着另一条写的：x.com/0xtrackmind/status/2106788767960850798