Jev 打分前置：先排序再开浏览器，三次运行共花 0.00083 美元

Klonzu @Klonzu原推发布于 2026.10.11

在 X 上打开原推

holy sh*t. staged GrokBot + Jev demo: an agent kills a dead retry loop in 4 seconds for $0.00040. stop paying a frontier model for cheap decisions make the agent rank before it opens anything "Rank every candidate cheaply, then open only the top 5 in the browser" Flights: 12 one-way fares Zurich to Heathrow ranked in 0.4 s for about $0.00002, SWISS 06:30 nonstop at $129 on top, booking off Posts: a pool of 24 scored on topic, reach and originality, only the top 5 opened, 4.1 s instead of 53.8 s Dead site: HTTP 503 three times with no state change, 11.2 s and $0.00168 burned without the gate, STOP_RETRY with 0 extra retries with it The big model only needs to weigh in at three critical moments: → before the browser opens: is this page in the top 5, or one of the 19 loads you can skip? → when a call fails twice: did anything change, or is the failure fingerprint identical (Δ = 0)? → before anything gets booked: did I confirm it, or is booking still off? Rank cheap. Open few. Stop on Δ = 0 This is the Jev engineering pattern I keep pointing at, one layer down: the mechanical calls (which fare wins, which page is worth a load, retry or stop) run as a scoring pass for fractions of a cent, $0.00083 across all three runs in this clip, so the expensive model only wakes up when there is a real choice Score first. Cap the browser. Cut the loop. - the full demo setup > run 1: http://flights.search ZRH → LHR, jev.score price · duration · stops, top 1 of 12, confidence 0.94 > run 2: posts.collect 24, jev.score topic · reach · originality, http://browser.open top 5 only > run 3: http.get 503, fingerprint_state compare previous, policy.retry STOP > cost per run: $0.00002, $0.00041, $0.00040 > fingerprint a3f9 c21e 7b04 9d10 matched, 2 wasted attempts avoided paste this prompt into Claude Code below: "Set up my Claude Code agent to make cheap decisions before expensive ones: 1. Before any browser or web fetch step: > Score every candidate URL on topic, evidence and duplicate risk from metadata only > Open only the top 5 and list the rest with their scores 2. Add a retry rule to CLAUDE.md: > On a failed call, record the status code and a short hash of the response > If the next failure has the same status and hash, stop and report instead of retrying 3. Add a PreToolUse hook that logs every retry with its attempt number 4. For anything that books, buys or sends: > Put it on ask in .claude/settings.json permissions Show every configuration diff first. Do not apply edits until confirmed"

这条是引用着另一条写的：x.com/completeskeptic/status/2108594987177021737