Jev in Practice: Browser Use and Agent Workflows

On this all-news episode, around the 31-minute mark, the hosts introduce Jev as a model that helps agents make fast decisions rather than write text, alongside segments on the revived Hyperloop, Trump's proposed AI Force, and open-source token share.

Cua released Cua-Bench-S1, a benchmark for decision models built for computer use that includes Jev, along with two checkpoints, Cua-S1-Nano-0.1 and Cua-S1-4B-0.1.

Lightfield's Chief AI Officer benchmarked TypeSafe's Jev on text understanding including customer-conversation tasks; with extended thinking off, Jev matched Opus 5 on accuracy at 40 to 150 times lower cost, and he outlined how Lightfield plans to use Jev in its own platform.

The author hooked Jev into RollerCoaster Tycoon 2 as the controlling brain, but it only built a few rides before getting stuck, showing it is not a plug-and-play game-playing model.

Using Jev to run a trust check at every step lets an agent representing a human gradually disclose more information about its principal as trust grows.

The author jokes that top San Francisco engineers are using Jev together with the fly brain to build code factories and ship faster.

The author is migrating the AI evals course repo built at Braintrust to PydanticAI, Logfire and pydantic-evals, and plans to explain where Jev fits in the end-to-end evals pipeline.

The author tested Jev and Opus 4.8 as supervisors (routers) in a multi-agent system built with LangChain.

The author used Jev to classify posts by context and then generated 1,001 individual icons and labels by category, noting it is not as cheap as people say.

An NVIDIA researcher discusses small models handling agent subtasks, with a Jev guide that classifies each email and lets code pick the next step, including a 50-message trial and human-review rules.

The author argues you shouldn't expect Jev to solve computer use by itself, but should use it to build primitives that agents and software can reuse.

Chroma CEO Jeff Huber explains how to use an LLM to narrow 10,000 candidate chunks down to the 20 that matter, with a guide to building that screening layer with Jev and testing it on 50 papers.