Jev wired into the explore step to pick the next element
The author uses Jev to walk a live app and choose the next element to interact with, moving routing, filtering and compaction decisions off language models for faster, cheaper classification.
JEV MADE A WHOLE CATEGORY OF PROMPT ENGINEERING POINTLESS IN A WEEK not the hard parts. the parts nobody should have been paying a frontier model for in the first place. routing. picking. filtering. deciding what survives a compaction. every one of those is a classifier we dressed up in a system prompt because that was the only tool on the desk. a model that cannot write one sentence does them 200x faster, 400x cheaper. we wired it into our explore step this week - it walks a live app and chooses the next element, no language model in the loop. it got faster and it got more boring, which is the point. the uncomfortable part: it is right more often and it can never tell you why. would you let something that cannot explain itself decide what your agent touches next?
This post quotes another post: x.com/me_barnyx/status/2101976764779999381