Jev in Practice: Browser Use and Agent Workflows

Of 1,100 calls, 1,003 were decisions, so the builder moved them to Jev and sent low-confidence cases to Kimi K3, dropping the monthly cost from $84.50 to $7.30.

The author argues that running Jev first to filter most input and letting the expensive model see only what survives costs about $0.06 a day for screening papers.

A Microsoft podcast episode covers Jev's idea of replacing generation with fast categorization, framed as a concept-level introduction rather than a hands-on guide.

The author places Jev in front of ChatGPT to classify and screen bug reports, so only well-evidenced reports worth handling enter the fix pipeline.

The author built an agent directory where Jev picks the model tier, chooses the next worker, and judges completion and risk, while Claude only generates text and every decision leaves a receipt.

COORDINATOR aggregates decision models including OpenAI, TypeSafe's Jev and Cloudflare's Clef to deliver split-second calls based on the user's context and preferences.

Opus splits tasks, Haiku executes them one by one, and Jev only checks each against its "done when" line and returns pass, fix or reject; nothing merges without Jev's pass, with a Claude Code config included.

The author trained step Jev to score each search and final Jev to read the whole trace and predict answer correctness (AUC 0.92), after an off-the-shelf hosted judge was gamed by the agent answering without searching and rated those wrong answers about 70% likely correct, for 0% accuracy.

Running Claude Code through a loop of one number, one editable file and a fixed time per run, targeting dollars per accepted ticket with Jev in the review, cut cost per ticket from 10.66 cents to 6.12 cents.

In a test by Lightfield's Chief AI Officer, Claude Opus wrote classifiers and TypeSafe's Jev ran them, and the agent-built classifiers matched or beat a conventionally fine-tuned model on all six tasks.

Across six classification tasks, classifiers built by Opus matched or beat fine-tuned RoBERTa on all six.

Open-source arc-cua replaces the screenshot-plus-LLM call on every step with decision models, supports Jev and the Decisions API, and aims to be a faster computer use layer.