Jev in Practice: Browser Use and Agent Workflows

A developer uses Jev to auto-select the agent, model, computer and folder: Fable plus Claude Code for major rewrites, and iOS changes run on one of his Macs.

AX now supports Jev in TypeScript, Python, Go, Rust, Java and C++, using ordinary ax signatures for boolean and class decisions or the native client for probabilities, rich criteria and scoring.

He had been using local models for the same task, which were slow and inaccurate, and switching to Jev just worked; he included a demo video.

The author gave an agent access to Jev through the Codex harness and tested whether a small, fast decision model can make SRE agents more reliable on 10 SREGym-Lite problems.

Box showed Jev classifying and routing incidents: it pulls the report from Box, builds state, asks whether it's customer-facing and how severe, then files it into Escalate, Monitor, or Review and sets metadata, with low-confidence answers sent to human review.

Union says Jev helps an agent decide what to do next while Union remembers what it already did after a machine crash, and the two would work well together.

A programmer says they favor Jev over open-ended LLMs because its constrained judgments are controllable and let them write small domain-specific languages on top.

A Turkish developer proposes embedding Jev in agent harnesses so a small decision model handles task classification, model routing, tool-call judgment, and checking whether work is actually complete instead of calling a large LLM for every judgment.

The author tests Jev's Choice, Score and Noul decision types, parallel sampling and RLCD confidence mechanism, and concludes the future is Jev plus LLMs rather than Jev replacing them.

The author walks through calling Jev's decisions endpoint on OpenRouter and tests it on ticket classification and Tetris moves, with a full game costing under half a cent.

The author tests Jev as model selector, skill selector, guardrail, memory compaction and multi-agent orchestrator, running a low-cost DeepSeek 4 Flash setup to see how cheaply Hermes-level capability can run.

After trying Jev, Eric Simons says these purpose-built, very cheap models will likely take some frontier share, and asks whether frontier labs will embrace the competition or face an innovator's dilemma.