Jev in Practice: Browser Use and Agent Workflows

A tutorial shows routing an agent's branch decisions to Jev through the Typesafe SDK using choice, score and noul question types, at $0.42 per 10,000 decisions.

Replaying 231 decisions with Open-Jev-27B on a single B300, routing and classification were nearly all correct at about 15 ms, date checks got 2 of 15 right, and the slowest decisions took 54% of the time while failing silently.

Routing 1,000 decision cycles through an Opus 5.5 + Jev harness drops the cost from $480 to $0.14, with Opus drafting plans, Jev scoring and abstaining, and Rust validating execution, plus a 6-stage runtime and latency/cost benchmarks.

The author sketched a router in about 15 minutes that picks a model based on the initial prompt, with DSPy handling the flow and Jev making the routing decision, and was surprised at how well it works.

The agent setup closed 312 tickets in a week versus 96 for 25 engineers, costing $2,055 a month in tokens against $343,750 in payroll, with code from Opus 5.5, tests from Dots, and Jev reviewing before anything is marked done.

A 21-line support agent uses Jev to sort 20,000 events down to 800 that need a reply, costing 8 cents for 18,700 events, and routes anything below 0.60 confidence to a human without ever sending on its own.

Hard filters handle the obvious criteria first, then Jev judges purchase intent, abandonment reasons and return risk, and Replit Atta analyzes and visualizes the enriched data.

About 100,000 signals from six data sources all pass a typed yes/no Jev judgment before any order is placed, with risk capped at 1% per trade.

Claude Code's Agent Teams split coding work across roles, while Jev handles mechanical actions like opening files and calling the CLI in 16ms without involving an LLM.

The Jev inference call adds about 100ms, keeping voice-to-voice latency well under the 1,500ms P95 target, with full code in the author's thread.

The author combines six Jev classifiers into one router for reranking, tool-call gating, citation checks, model routing, injection detection and page classification, claiming it is 444x cheaper than the frontier model.

The architecture has the LLM write, Jev recommend and the app check permissions, so an agent whose approval was revoked cannot perform irreversible actions.