Jev in Practice: Browser Use and Agent Workflows

The game ORBRYN is wiring Jev into world agents that make fast decisions, trade, gather, fight and transact onchain alongside real players, with a demo promised later.

A developer finds Jev suited to limited-solution-space, low-latency decisions like computer use and game AI, but says tool-use parameter reasoning still needs an LLM, and is building poker and Pokemon VGC AI demos.

Jev combines price forecasts, technical indicators, time-series data and news feeds to find trading opportunities in real time.

The author argues Jev is essentially a classifier that keeps the LLM's cross-task generalization but drops language generation, making it suited to routing, judging and verification in agents.

Developer thekitze built a smart calculator notebook using Jev, which he says can calculate anything, with a demo video.

A repost describes using Jev to race a game built in Combos, combining fast decisions with a prompt-made racer.

LangChain's Sydney Runkle uses Jev for model routing and pre-tool risk checks, and the post rounds up community cases in browser automation and trading agents.

A video post observes that last week people were building 3D games with Astra, and this week they are all writing trading bots with Jev.

The author open-sourced jev-visual and recorded a gesture recognition demo that runs locally on a 16GB M4 MacBook Pro with about 200ms per recognition.

A Maxfusion tool uses Jev to rate each shot of Meta video ads, making 50 judgments in 1.52 seconds for $0.00028 and processing 450+ ads in under 3 minutes.

Cua's jev-use splits Computer Use into a Jev decision layer and a Cua Driver that observes and executes, and in a 2048 test it ran 5x faster and about 1,000x cheaper than Astra.

Using typesafe-sdk, Jev made structured judgments on 724 live ads, returning fields like hook, CTA and awareness stage at a median of 216 ms per ad and a total cost of 9 cents.