Jev in Practice: Browser Use and Agent Workflows

The author open-sourced System1Code, a coding agent that uses Jev for many small decisions and leaves only the hard parts to a frontier model; on the same bug-fix task, all tests passed.

A self-deprecating joke about building an auto-reply agent with Jev, after which the wife asks who Jason is.

A discussion of three practical directions for the System One decision model Jev: continuous evals and monitoring, custom verifiers for long-horizon agents, and a recursively self-improving agent harness.

Someone used Jev in a simulated car for steering, lane changes and merging decisions, building the loop in under an hour; the model returns actions with probabilities and a confidence score rather than text.

A demo shows how GEPA can take an off-the-shelf model like Jev, trained only on synthetic data, and tune it to your exact decision criteria by adjusting the prompt and labeling a few examples.

At Claude Build Day, a demo shows Jev deciding which Claude model and how much reasoning effort a task should get before the main model is woken up.

After an hour of testing, the author describes giving Jev context plus a question and getting back a structured choice, a 0-100 score or a probability, for uses like agent routing, output checks, workflow triggers and human escalation.

The open-source CLI jev-align uses GEPA to teach Jev what good and bad examples look like, addressing the need to calibrate a foundation model to your own decision criteria.

The author used Jev to help optimize several parts of PBIR-CLI, then generated this report with Claude Code using Opus Low in about 20 minutes.

A video of an AI agent workflow running, with the poster saying they want to try it on Jev to see how it performs.

TypeSafe, founded by former OpenAI researcher Diogo Almeida, introduced Jev, a model that returns predefined decisions with probabilities and confidence estimates for software to use directly rather than generating text.

ClickHouse's aiFilter uses Jev to semantically filter and score agent sessions at about 50 rows per request and 300ms, treating Jev as a database primitive.