Jev in Practice: Browser Use and Agent Workflows

The author spent a morning testing Jev, having GrokBot control Workbuddy and Codex control NetLogo, and found it interesting and, crucially, cheap.

The author open-sourced the Jev-cu repo, which lets Codex operate various apps smoothly once the skill is installed and the Jev API key is configured, with bigger speedups on tasks that require frequent judgment.

The author argues Jev handles low-entropy tasks like classification, routing and scoring with one-pass typed decisions, speeding up multi-step workflows such as browser use by over 100x, and that the future splits work between cheap System 1 models and frontier LLMs.

The instructor showed several prototypes built with Jev in an AI cohort, stressing the goal of building a mental model rather than just wiring up a solution.

The author set up a small chat harness to interact with Jev, describing it as like talking to Rocky from Project Hail Mary: when it is certain, it is very certain, and tagged typesafeai.

A developer used Jev to read 36 fictional dating profiles, give a verdict on each with the evidence behind it, such as "career talk is fine, asking me to invest isn't," turning vague preferences into explainable decisions.

The open-source JevRouter project lets agents like Codex and Claude Code use Jev without rewriting orchestration, claiming 38% vs 24% tool hit rate on the first 5 Toolathlon steps, about 5.5x faster and roughly 7x cheaper.

Skeptical of online demos showing Jev playing games perfectly in real time, the author tries to replicate one and finds the hard part is the missing game state and the bot code that actually plays.

The author argues there is no need to build a custom ML model for one-off tasks: when the large-scale classification task changes every day, having an agent call Jev is simpler.

The author posed Jev's self-identity as a discrete-choice question and Jev answered that it is OpenAI; he confirms the team never taught the model to say it is typesafe/jev.

The author asked Jev about 3,000 yes/no questions on Zuckerberg's Senate testimony, completed in 3.5 seconds for one cent, and colored the page by the answers; users can also type their own questions to reread the full text.

An open-source project lets the Jev model execute phone operation commands directly on Mobilerun cloud devices without an ADB connection.