Jev in Practice: Browser Use and Agent Workflows

BohuTANG says many agent tasks are classification problems that take the long P(token) route, while Jev's P(option|state) gives option probabilities directly, a promising path.

The author argues Jev's difference is not new capability but being a general classifier that is fast and cheap, since slow demos could be done with existing models.

Jev is now callable on Cloudflare, returning only a decision and probability in yes/no, choice, or graded-score form, with low-confidence cases routed to a human in code.

A 16-minute video covers community uses of Jev such as browser control, email sorting, and code review, plus demos of rabbit-adventure route decisions and alien entry judgments, with 279 calls costing 2 cents.

A breakdown of Jev's three input types, yes/no, single choice, and scoring, and how Browser Use mainly relies on Choice to pick a click target from accessibility tree nodes.

The author says they are testing Jev for agent evals and waiting to see how it turns out, with a screenshot of the test process.

A usage sketch that chains Hermes Agent, the Jev decision model and Computer Use into a single agent workflow.

An early prototype uses Jev to drive state machines, which the author says could cut seconds off a typical agent run and suits neurosymbolic programs well.

mizchi shared a research report on chaos-testing browser scenarios with Jev in the jev-playground repository.

In word-count classification tests, injected instructions pushed Jev's output probability from 0% to 43%, leading the developer to hold off on adversarial uses like resume screening and invoice approval.

Speko now routes through Jev, returning typed decisions with calibrated confidence each turn so agents can branch mid-call on intent or handoff without parsing text.

A user describes joining the TypeSafe waitlist and answering a short questionnaire to get a Jev invite in roughly half a day, noting Jev returns structured decisions with confidence estimates.