Jev in Practice: Browser Use and Agent Workflows

The author lists Jev-based products such as one that picks what to paste and a Jev input method, and argues that AI making choices for you based on context is finally taking shape.

A roundup of NYC tech and AI events this week includes a 9/23 talk on Jev, open weights and coding agents, and a 9/24 Jev demo night at betaworks.

The author argues that instead of letting the main LLM choose skills, a classifier model like Jev should decide which Agent Skill to invoke, keeping the LLM context clean even with hundreds of skills.

The author argues against bolting Jev on directly: first trace the decisions an agent repeats fifty times an hour, then route those through Jev with model fallback and verification.

A Chinese post shows a case of recreating a video with Jev and includes a demo link.

The author says they built a generic classifier before Jev became popular, sampling a single token from gpt-4.1 and using logprobs to assign probabilities to each option, and shares the aixcc project code.

A video demo shows Jev replacing generative classification with a single forward-pass score, with the author noting that sometimes you just need a smart if statement.

The author shares how they use Jev to score work in content, recruiting and video editing, noting that a high score is not proof the work is good and some checks still need human judgment.

The author replaced their prompt tracking classifier with Jev, cutting runtime per run from 1,765s to 6.3s and cost from $0.944 to $0.002.

The author built six 3D residents with their own homes, jobs, needs and friendships, and Jev decides whether they grab food or seek shelter, letting users inject a situation and watch the small world react.

A list of eight open-source repositories that use Jev as a decision layer, from browser control and faster search to context compaction for Claude Code.

The author argues that Jev's classification and scoring abilities exist in other large models too, and that its real value is faster, cheaper and more stable structured output, while asking how the accuracy of its probability judgments should be measured.