Jev in Practice: Browser Use and Agent Workflows

The author tagged all 1,284 tool calls in one session and found 938 were simple yes/no decisions; routing those to Jev lowered the bill from $64.77 to $26.16 at 38ms per decision.

The author says Jev can neither code nor write words, and shows what they built with it yesterday.

LangChain tested Jev on accuracy, repeatability, latency and cost to see whether System One models could offer a new approach to agent evaluation.

After a day of hundreds of trials integrating Jev into Codex to help make browser-use decisions, the author has spent $0.00725 in total.

A roundup of eight open-source projects that use Jev as a decision layer, covering browser control, search, context compaction and tool calling for coding agents.

A developer used Jev to add voice commands to aloud.rest, including setting a timer, asking the facilitator to speak slower or faster, ending a session and saying "show me the orb".

In a talk from about two months ago, Jev's founder argues that earlier models were limited by RLHF and built for human-in-the-loop use, while Jev targets true automation as the next era of AI and software.

LangChain is holding a livestream on Tuesday with Typesafe about building a harness with Jev.

A one-line take that Jev delivers the general-purpose, classifier-style branching for agents that the author had always wanted.

The author wired Jev into Codex for computer and browser use, cutting subagent handoff and warm-up time, and says the loop now runs in a working state.

The terminal uses Jev to score each crypto tweet as bullish or bearish, combines that with funding rates, RSI and volume, and outputs entry, stop and target cards; the code is open source.

The team uses Jev in its Codex workflow to recommend the right agent, context and review level for each task, with Astra leading the build and Sol helping execute.