Jev in Practice: Browser Use and Agent Workflows

Running one million decisions through Claude Opus 5 cost $176,100, while the same workflow on Treg orchestration plus the Jev decision layer costs $400.

On the same slide deck and brand guidelines, the Jev path took 16 seconds and 3 cents while Claude ran 10 self-checking steps and burned 70,000 tokens, with the fix being Jev choosing layouts from a fixed set and the LLM only filling content.

LangChain showed Jev judging task difficulty, blocking risky tool calls at runtime, and grading agent answers against rubrics in online evals, with Jev answering PII detection instantly at 98% accuracy.

Outreach is broken into a series of decisions, and Jev scores each one, from whether a lead is the right person to which message will get a reply, then locks the best message per person and builds the campaigns.

Simple map commands like zooming to a layer or switching to a dark basemap are routed by TypeSafe's Jev instead of a full LLM round trip, cutting multi-second execution to sub-second, while complex requests still go to the full model.

FrontiersMind fine-tuned a 0.6B decision model on its from-scratch Lumma-0.6B-Base that takes a document and typed questions and returns a probability for every answer in one forward pass, generating no text, for routing, triage and moderation.

A demo has Grok 4.7 do the thinking and Jev only answer yes/no, pick one option or give a score, asking 6 questions each of 3,412 leads to complete 20,472 decisions in 15.7 seconds at a cost of $0.41.

The OpenClaw author explains how decision models are being integrated into OpenClaw core and plugins, and links a blog post on how to contribute.

A Stanford team replaced LLM interpretation of every result with a Jev-based decision layer that routes and scores records, making production up to 193x faster and 444x cheaper.

Pydantic says Jev powers Pydantic AI agent calls to return faster, continuing Pydantic's type-safe approach.

A scout, architect, builder and operator agent, each pinned to its own model version, run in overlapping lanes via Jev handoffs, cutting 80 serial minutes to 60 at a cost of $4.55.

A leaked Jev engineering file describes eight agents (planner, researcher, coder, critic and others) that exchange only typed decisions, each returning a result, explanation, confidence and next steps.