Jev in Practice: Browser Use and Agent Workflows

A developer used Jev to rebuild a website's internal link map, processing 586 pages in 45 seconds for $0.21, with an approach where the LLM plans, Jev routes and scores, and tools execute, cutting calls to expensive models.

After using Jev, the author argues for a model whose sole job is to find ambiguity, ask clarifying questions, and name the decisions a human or agent needs to make, and hopes typesafeai is working on it.

Across 16,500 predictions on 8 real datasets covering loan defaults, bankruptcies, and medical diagnoses, Jev failed calibration on 7 of 8 and systematically overstated the probability of bad outcomes by roughly two times.

The author suggests using Jev only for the decision steps in an agent: picking the cheap path, checking a step before it runs, checking results before they go out, and trimming long runs, recommending starting with the last one.

In a 10-minute breakdown, Steve Sewell explains how picking tools and skills per prompt keeps irrelevant instructions out of an agent's context, and how to build a Jev router that sends uncertain decisions for review.

A developer uses Jev for podcast prep to decide which questions to ask, and asks others to share their own Jev demos.

Teknium breaks down "Jev compaction": it only removes tool-call tokens, shrinking by half each round until nothing is left to compact, which is why Hermes Agent has rejected the PR every time.

Bespoke Labs built Nimble, a 9B scoring-based decision model, from Qwen3.5-9B with LoRA and 2,676 examples in one day; it scores 90.12% on a 324-example test, below Jev's 93.21%.

The author cancelled their ChatGPT subscription after a week with OpenJev, an open-source tool that needs no API key, keeps data on the local machine, returns decisions in milliseconds, and can batch-classify emails, comments, and job listings.

Parker's co-founder says the entire engineering team is reimplementing the app's core flows with Jev, so users will benefit from the model indirectly.

The author built an email app with Jev that prioritizes the inbox live by importance instead of reverse chronological order, coming soon to avec.

The post proposes a two-stage computer use flow where an LLM generates the state and a multimodal Jev makes the judgment, calling it the start of low-cost deterministic intelligence.