Jev in Practice: Browser Use and Agent Workflows

TypeSafe published the typesafe-ai/skills repo, installable with one npx command, so coding agents can use Jev to answer repeated decisions like routing, validation and whether human review is needed, cutting expensive frontier-model calls.

The author built a CLI and Python library that generates numbers with secrets and asks Jev whether each looks random enough, with rejections coming with a reason, plus install command and GitHub repo.

The author built a CLI experiment where Jev reviews code changes before a commit: it classified all eight constructed cases correctly in 359 ms and distinguished a harmless log line mentioning a token from an actual session token leak.

The post argues Jev is not a replacement for coding models but a decision layer that handles small choices like retries, tool routing, risk scoring and whether human review is needed, so expensive models are only called when real reasoning is required.

The author says Jev is being widely adopted because every decision returns a typed structure, and collects 11 concrete usage examples.

DeepForAI says it has spent nearly four months training a Jev-like decision model focused on Arabic and local deployment, may open-source it within weeks, and links an article on building non-autoregressive decision models.

Jev is TypeSafe's System One model that takes context and returns typed decisions such as choices, scores and probabilities, suited to classification, routing, scoring, guardrails and real-time product decisions.

Replacing per-cycle re-reasoning with a 10-step Jev loop of state capture, ranking, scoring, selection, routing, execution and verification reduces repeated reasoning and ambiguous routing.

A developer teases that combining Jev with agent-browser enables some wild things, sharing one example for those who want to see browser agent decisions in practice.

The author argues Jev's opportunity isn't a cheaper model but putting it in front of expensive human queues like lead scoring, ticket triage, application ranking, and transcript clipping.

The open-source Stanley project pairs small-model judgments with deterministic workflows so coding tasks stay bounded, inspectable, and able to improve over time.

Ex-PolyMarket researcher Henvy released a meme trading terminal called NERVE, with seven autonomous node AIs using JEV and ASTRA in a trading architecture, now open-sourced on GitHub.