Decision type · agent loop
Where Jev fits in an agent loop
A chat model in an agent loop does three different jobs with one mechanism. Jev replaces exactly one of them: the bounded decision between alternatives the loop already knows about.
01Three jobs, one of them is a decision
Deciding which tool to call, whether a draft violates a rule, and what to write are different problems. Only the first two are enumerable, and those are the ones that get faster when the mechanism stops being generation.
prompt ──▶ [ Jev: which skill / tool / params ] ──▶ agent executes
│
uncertain ────────┴──▶ chat model (the tail only)02Routing the uncertain tail
The pipeline that shows up repeatedly sends only the low-confidence cases to a larger model. The economics depend entirely on how big that tail is, which is why threshold choice deserves its own measurement.
03Failure modes
SKELETON — wave-1 prose pending: a wrong routing decision is invisible until output degrades, so routing needs its own eval set.
Case evidence
4 of 336 indexed- @Mnilax
Jev inserted into the routing layer of an existing build.
View on XAuthor-reported
same task 37 minutes sooner, zero malformed responses to retry
Not verified by us
- @johnyeo_
Slack agent classifies skill, tool and params before handing off.
View on XAuthor-reported
agent 2× faster
Not verified by us
- @nutlope
Two-stage pipeline: Jev classifies, uncertain cases escalate to a chat model.
View on XAuthor-reported
100 emails in 1.42s, 96/100 correct, ~$0.07
Not verified by us
- @GNUmanth
Pre-flight gate using four typed primitives before running an install script.
View on XAuthor-reported
~100ms gate
Not verified by us
Metrics are as posted by the author, captured 2026-09-20. We index and link; we do not verify, and we do not republish the post or its media.
Related