Jev in Practice: Browser Use and Agent Workflows

Someone open-sourced fast-jev-compaction, which uses Jev to compact context for agents, with a GitHub link.

The author and a collaborator introduce Jev, TypeSafe's newly released classification model, said to be up to 200-400x faster and cheaper than comparable LLMs on classification tasks, and preview a tutorial on integrating it into a harness.

Browser Use founder Gregor Zunic ran a real flight search with Jev, finishing in 7 seconds at $0.0039 per run using a dynamic action space, token-free structured output, and a large/small model fallback.

A short list of uses to try with Jev right away: LLM-as-a-Judge evaluation, routing for agent harnesses, and scaling agent orchestration.

Jev is now available on Cloudflare AI Gateway and can be called directly via env.AI.run, with documentation published.

A shared Every vibe check article explains how the Jev model works and describes it judging everything the author had written in 0.7 seconds.

An argument that agents may split into two models: an LLM deciding what to solve and how to explore, and Jev deciding each concrete step based on probability, latency and reliability.

A long post breaking down when Jev applies: limited actions and goals, clear feedback, observable state and high-frequency decisions, plus a value formula.

The author says auto mode had been turned off because classification was slow, and is being turned back on now that Jev is in use.

OpenRouter lists Jev at $0.042 per million input tokens with free output, 70-500ms end-to-end, and up to 190x faster and 440x cheaper than an LLM on TypeSafe's published workflows.

The author built a proxy that asks Jev which parts of the conversation history are still needed and drops the rest before the request goes out, getting 8/8 exact answers on the real APIs with up to 50% less context usage.

After a day of testing, the author lists four internal use cases for Jev: task-difficulty routing, orchestrator-to-subagent communication, social content signal classification, and email triage, concluding it fits best where outcomes are deterministic but the workflow is intelligent.