Jev in Practice: Browser Use and Agent Workflows

Ollama now supports running Jev-style decision models entirely locally, and the author says the in-house demos are addictive to play with.

The team integrated Jev into the /last30days skill, scanning 10x more posts at lower cost and raising sandbox launch recall from 5/13 to 10/13.

The author argues that once agents enter real loops, the bottleneck is no longer how smart the model is but the need for cheap, instant choices, which is why models like Jev and OpenAI's Decisions API are emerging.

A developer says Jev-like decision models perform well at playing real-time games, but development also surfaced many problems; a demo is included to try.

Qwen3.8-27B was plugged into a live game loop via SGLang's /v1/decisions and /v1/systemone endpoints to score actions, beating the Elite Four with decisions under 100ms each.

The project argues DeFi's problem is decisions rather than yield, and uses autonomous agents plus JEV to weigh thousands of pools and decide where capital should go.

The author shares a prompt to paste into Claude Code that they say builds an Opus 5.5 trading bot making decisions with Jev.

Inspired by Jev, the author built a 29MB System One decision model that runs directly in the browser and performs reasonably on familiar tasks, though generalization is still limited.

A Reddit user connected Jev to on-chain data streams to spot arbitrage opportunities for free, then hands the good ones to an LLM to buy, bridge and sell, at roughly the cost of a basic subscription.

Given a homepage URL, jev-seo crawls the whole site in about a minute and produces a prioritized fix list against 52 rules from Google's search documentation, with Jev judging each page's topic, content specificity and keyword overlap.

The author breaks down a four-part AI trading bot: Opus 5.5 runs backtests and designs the strategy, Jev reads live market data and returns probabilities, and rules plus the exchange handle execution.

SGLang turned Qwen3.8-27B into a multimodal decision model that beat Pokémon FireRed's elite four and champion with sub-100 ms decisions, and added /v1/decisions and /v1/systemone endpoints.