Jev in Practice: Browser Use and Agent Workflows

The author proposes a new AI primitive between a model call and an agent loop: pass in a prompt, typed questions and thresholds, and the reasoning model loops until the Jev decision model says the answers meet them.

Lovart's MCP is now live on Codex, letting users batch-produce brand creatives and ask which ones will actually perform, powered by Jev.

ScienceBuddy is a free workspace for scientific agents offering GPT-6 at no cost, built on a Jev framework with an inner loop that refines the agent harness and an outer loop that trains the model via rubric-guided RL.

Eachlabs built four tools on Jev that score 907 models, 95 art direction skills and 24 slop tells before rendering, aiming to cut wasted renders and save time and budget.

PhAILabs released ScienceBuddy, a free scientific agent workspace that uses GPT-6 and fuses the JEV framework, refining the agent harness in an inner loop and training the model with rubric-guided RL in an outer loop.

A developer open-sourced a unified agent harness layer where System One (powered by Jev) and 13 other harnesses including Codex and Claude Code share one protocol, and it runs locally for free.

The post lists five real uses: blocking destructive Claude Code commands, sub-second fraud checks on invoices and refunds, overnight inbox sorting, scoring 500 applications in a minute, and tagging comments and tickets.

The article argues for Jev as the next-step decision layer for agents, deciding in milliseconds at near-zero cost, and says setting it up right can cut most of an agent bill without hurting quality.

The author ran the same 500,000 classification, routing and scoring decisions on Opus 5.5, GPT-6 Astra and Jev; Jev cost 21 cents and was more accurate, and the post outlines how to move decisions from generative models to Jev.

A third-party paper tested Jev on LoCoMo for memory write, retrieval and stop decisions: memory build took 158 seconds versus 1,044, query time fell from 1.47 to 0.93 seconds, and the score was 0.777 versus 0.700.

The author breaks down six steps for using Jev to speed up an agent: structured state, isolating repeated decisions, a Jev decision layer, smart routing, verification and memory updates.

Jev sits at the front of the email queue and only makes judgments without writing text, handling 1,700 emails for 18 cents with a confidence score on every step.