Jev in Practice: Browser Use and Agent Workflows

jarrodwatts has open-sourced a Jev trading bot demo that can be tried online or inspected in the GitHub repository.

A demo shows Jev reading dozens of structured trading signals and producing intelligent decisions, which the poster says makes a trading system actually alive.

Vercel's fx harness replaced GPT-5.6-luna with Jev for judging whether commands are harmful, making the check 18x faster, cheaper and slightly more accurate.

The author built a semantic lens on Jev and LanceDB that filters and explores data in near real time without predefined features, and shared an online demo.

A developer vibe coded a parallel searcher where a BFS queue discovers new pages and Jev consumes and reranks them in parallel, with plans to add a local LLM for human-readable output.

The project asked five yes/no questions per sentence for both candidates, totaling 1,191 Jev calls and 1.18M tokens at a 415 ms median latency and $0.0497 total cost.

A developer lists decisions Jev could handle, including harness and repo selection, fast computer use without Astra, whether to reply in a Slack thread, and sandbox sizing.

A 12B probabilistic model generates price paths and indicator evidence, which Jev reviews to output LONG, SHORT or NO TRADE with entry, stop loss, take profit and invalidation rules.

Page elements are turned into an ordered list of options for Jev to choose from, with a small model called only when text input is needed, completing a Google Flights search in about 7.1 seconds.

A proposed System One architecture where the LLM only does token prediction while Jev decides when a tool call is needed and injects the results back into the token stream.

The bot replaces regex filters with Jev decisions, adding four-stage progressive escalation, dual audit logging, two-step admin confirmation and one-click pardons that teach the model to avoid repeat false flags.

The author sees Jev not as an LLM replacement but as a partner handling the judgment step, solving slow LLM decisions while being fast, cheap, and free of format hallucinations.