Jev in Practice: Browser Use and Agent Workflows

Proposes splitting bug fixing so Claude plans and writes code, a graph structure separates building from review, and Jev verifies evidence before approval, leaving reusable checkpoints for repeat fixes.

The author compresses signals from 49 micro-agents into a unified state, hands it to Jev 1.13 for a bounded buy/sell/hold choice, then passes it through a risk gate.

Pulling small internal decisions like whether to call Claude out of the large model into a separate decision layer cut the bill by 87% with almost no change in product behavior.

The JevBench author announces expanded evaluations covering real-world Jev-class use cases such as LLM auto-routing, retrieval augmentation, and content moderation.

jes is an open-source project that uses decision models like Jev to guard every step of an agent's trajectory: prompts, tool calls, tool results, skills and responses; the video was made with Opus 5.5.

TypeSafe's Allie Laabs joins The Merge to discuss Jev decision models and the habit coding agents still need to learn: asking questions in parallel.

Developer Timidan_x built Tern at the first SERV Hackathon, using SERV to reason over untrusted information, a Shadow Agent to review the reasoning trace, and JEV to score it before moving forward.

The author used Jev for sentiment analysis in Canonry, scoring 2,304 AI answers for $0.37 in about a minute, which they call 90%+ cheaper and faster than conventional approaches.

A short reply to bytebytego arguing that Jev is not the best choice as a model routing layer, with an image illustrating the point.

Jev compresses unstructured text into 38 bounded questions, 29 graded signals and 67 numeric columns, cutting model RMSE from 3.09 to 1.77.

The setup uses You.com for search, Jev for typed judgments and Qwen for proposals, integrated via MCP, at about $0.02 per sweep.

The author calls the combination underrated because of its division of responsibility: Jev stays in the decision loop while Opus 5.5 stays out, making the system easier to audit, and lists 10 architectural rules.