Jev in Practice: Browser Use and Agent Workflows

The proposal uses Jev for lightweight routing decisions under 100ms and leaves heavy reasoning to a large model, avoiding the cost of sending every micro-task to an LLM.

An open-source project uses Jev as a browser agent to run a Zürich to London flight search in about 7 seconds at a cost of $0.0039.

The piece argues for splitting agents into three layers where LLMs write, Jev decides and code executes, replacing free-form output with Choice, Score and Noul typed outputs, and includes batch-call benchmark data.

The author argues that Jev itself may be forgotten, but its idea of not using LLMs for control-flow decisions and instead using a small model for a single forward-pass score will be adopted across the industry.

The author built a semantic log filtering experiment with Typesafe Jev, including a CLI so an agent can use it, plus a TUI wrapped around it.

Without a solver library or teaching cube algorithms, code generates legal steps for the current stage and Jev only judges which of 12 is best, solving the cube in 17 decisions.

Developers discuss how Jev differs from Outlines and llguidance, asking whether its type safety is intrinsic to the model or enforced by the inference layer, and call for technical details.

The team spent the weekend pushing Nym's agent as far as it could go on Jev, and a blog post covers the benchmarks and performance characteristics, aimed at making agents faster, cheaper and more reliable.

Maryam Miradi walks through six steps on how Jev differs from chat models: it returns typed decisions with explicit probabilities, letting confidence drive run, review or escalate workflows.

A demo uses Jev to assign each line a drop probability, keeping the original error text and removing irrelevant lines, cutting tokens from 156,000 to 62,000 without rewriting content.

The author has the agent call Jev before each action, scoring it on six factors such as reversibility, budget and data backing; any red flag pauses the action and pings the author, at about $0.001 per decision.

A trader gave Jev 600 trade setups, market context and entry rules; it evaluated every setup in 47 seconds for $0.07, flagging those that failed the criteria, and can also score signals and detect conflicting indicators.