Jev in Practice: Browser Use and Agent Workflows

Another Jev demo video that forces every dog into wolf, pig or rat and every meal into soup, salad or sandwich, showing forced structured choices.

The author explains why Jev took off and lists 10 Jev projects in actual use, covering browser agents, desktop automation, code review and model routing.

The open-source trading agent AgentOS adds Jev as a third Pilot Router strategy: send state plus typed questions and get a decision with a confidence score in about 0.9 seconds.

After integrating Jev, Spawn makes 10 decisions per second, and creators already use it for NPCs, bosses and puzzles that fight back, generated directly through Savi.

Jev first filters a large pool of candidates before handing them to GrokBot, processing 3,412 people in 15.7 seconds for $0.41.

The author built two versions of a Jev workflow that classifies alerts full of same-name company noise into real signals, with the automated version pushing only useful mentions to Slack.

Three ways to install Jev into coding agents: the official Plugin/Skill, npx skills add, or letting the agent install it itself, so the agent does the work and Jev makes the judgments.

The author argues that forcing generative LLMs into discriminative tasks like classification and extraction is costly and unreliable, and that Jev brings discriminative models back into focus.

The author says Jev is small, does no reasoning and handles non-English input poorly, making it fine for suggesting a project in Todoist but unsuitable for critical decisions like auto-assigning one.

New API now supports Jev, turning natural language into structured decisions via the TypeSafe plugin while keeping the existing SDK.

TypeSafe's decision model Jev is now available on VectorAxis, sharing the same gateway as LLMs; migrating just means pointing the TypeSafe SDK's base_url at it.

Convai released Laya, an open-source non-autoregressive decision model that beats Jev on most benchmarks at 6-8x the speed and runs locally, though Jev still leads on 50+ label spaces while Laya defaults to about 20 options.