Routing agent loops through Jev to skip unnecessary model calls
The author shows routing judgments like whether it is safe, whether the task is done, and which tool comes next to Jev, which returns Choice, Noul or Score in 250ms, with cost and accuracy figures from a 100-email fraud pipeline.
Every agent loop burns money on calls that never needed a model. Route this. Is it safe. Are we done. Which tool next. Jev answers those in 250ms without writing a single token. Not because it's faster. Because it can't write at all. Here's what that changes: → Jev returns three types: Choice, Noul, Score. Up to 255 options, a probability between 0 and 1, or a level on an ordered scale. No free text. No parse step. No malformed JSON. → One request, many questions. Send thirteen judgments against the same state in a single round trip. TypeSafe reports 12.2x cheaper and roughly 10x faster than thirteen separate calls. → Confidence is the product. Pin jev-1.13.0, label 200 of your own examples, bin the predictions. The 0.9 threshold in every tutorial is a worked example. You read yours off your own curve. One team ran 100 emails through it, half fraud. Jev finished all 100 in 1.42 seconds. The 31 below 95% confidence went to a frontier model. Whole pipeline: 16 seconds, 96 correct, $0.07 total. The Jev share was a third of a cent. Three questions per turn across 200 Claude Code turns costs $3.02 a month. The same decisions through Sonnet 5: $229.50. The real shift isn't the price drop. It's that you can finally route the work you never wanted a model doing in the first place. 👇
This post quotes another post: x.com/0xcodio/status/2104195432209387524