Jev in Practice: Browser Use and Agent Workflows

A case study says Jev handles forks like picking which notes to load, delegating trivial steps, choosing recovery paths and ordering checks, while Opus only does hard reasoning, saving about $106 a day.

The story of Diogo Almeida and TypeSafe releasing Jev, a model that only picks from preset answers and returns a probability, far faster and cheaper than chat models.

The author shows putting camera frames into Jev's state and using Choice, Noul, Score and Box in parallel to judge gestures, whether something is present, and where it is.

A Polish developer releases basal 2.1 built on Jev's approach, adding act, tool, classify, verify, extract and compress, with open local weights.

A developer uses Jev to judge whether each iteration of an agent loop is making good progress, and says the early experiment looks promising.

The post links Atlas for spatial understanding, Jev for multiple-choice fast decisions, and Strata for running models locally, with a comparison of 0.1 seconds versus 8 seconds for judging whether an invoice is fake.

Two people won a five-hour hackathon with a navigator built on Jev that provides automatic steering for their agent harness.

On TermiX, the author joined a $1 Jev FC session as player one, left for food while the agent handled player two, and found it felt like watching a teammate learn the game.

A post argues for building durable System One state machine workflows with Effect and Jev, without giving any implementation or example.

In a comparison, Jev took 316ms per decision and fell behind the falling blocks into Game Over, while a local small model at 43ms played smoothly on a 16GB MacBook Air, concluding latency matters more than intelligence in real-time scenarios.

The author used the $1 JEV FC buddy to let an agent take over the second controller, with the game continuing while they stepped away, and tested the experience through an escrow-based payment flow.

During an Agent Loop livestream, the author tested Jev's new multimodal image input by building a small homemade bench that detects baldness.