Jev in Practice: Browser Use and Agent Workflows

OpenCode's changelog shows Zen now supports TypeSafe AI Jev (jev-1.13 and a free version) for parallel yes/no, multiple-choice and rubric questions, and adds models including Grok 4.7.

The author built a small tool with Jev for bug bounty hunters that checks whether a finding is a duplicate before submission, though it currently returns "probably" for everything; a copy is available on request.

A long-form article argues that Jev fills the missing decision layer in AI agent architecture and discusses its place in agent workflows.

On a 16GB MacBook Air, the locally run Laya System 1 model made decisions in 40–50 ms while cloud-based Jev sat above 300 ms, giving the local model the edge in live gameplay.

A Thai-language introduction to Jev, a System One decision model that only picks from given options, scores them or returns probabilities, suited to classification, routing and agent tool selection.

The author's own test finds Laya is fast because it runs locally with no network latency, while cloud-based Jev is slower but more accurate, so the two fit different use cases.

The Tinker team says an open LLM is already a probabilistic classifier, and a $5, 10-minute fine-tune lets it return discrete-option probabilities like Jev.

The author says that after building Jev into their marketing workflow, LLM call costs dropped to roughly 1/25 of the original.

The author says they built a new batch of Jev tools and used Opus 5.5 to generate an introduction video, with a link to the original post.

The author built JevSearch, which takes the top 25 results from Browserbase search and has Jev score them against selection criteria to return the top 5, often picking more relevant URLs outside the initial top five.

A weekend open-source demo uses Jev to re-rank SERP results: across 50 questions, 74% preferred the short list and 86% had the answer in the top three, cutting context noise and LLM cost.

The author lists 10 pitfalls from the Opus 5.5 migration docs, focusing on setting effort per message in Jev and fixing speed on the first turn to avoid cache misses and 400 errors.