Jev in Practice: Browser Use and Agent Workflows

On Venice, public agents can receive DMs: Jev reads each incoming message (such as "a top-10 backer sent gm") and decides whether to allow or skip it before the LLM replies.

The author shows how, after Hermes Agent claims a task is done, the evidence is handed to Jev for verification, and shares a reproducible prompt for testing your own Hermes Agent.

An introduction to Jev, released by TypeSafe on September 15: it takes a state and returns typed decisions (Choice, Score, Noul) with probabilities, at 70–500ms latency and $0.042 per million input tokens with free output.

The author says Jev lets an agent rate experiences by how deeply they resonate, triggering memory nodes that link into a graph, a key piece of the Mnemos cognitive engine.

A practical tip: create an eval, run Jev on your prompt to check the eval, then iterate based on the report, which the author calls the best way to improve your product with Jev.

An open-source PR agent built with Jev scans 384 news stories in 24.9 seconds and picks the best ones for 15 brands to jump on, at a cost of $0.19, with 30+ skills included.

The author compares their own image-generation results with Jev's demos, noting that the simulated road and obstacle course lack real sensor noise and failure modes.

The post cites official figures: Jev makes a decision in 0.114 seconds versus 8.566 seconds for an LLM, costs $42 per billion input tokens, returns typed decisions with confidence scores, and has an open waitlist.

The author argues that System One models like Jev will mainly make agents dozens of times faster with far fewer errors, and asks how Anthropic and OpenAI will respond.

A 20-minute video walks through three Jev demos: selecting DOM elements by voice in 150ms, scoring away 90% of redundant memory, and batch-predicting YouTube hit probability.

Jev, which only makes decisions instead of writing sentences, spread quickly on Vercel AI Gateway by returning typed answers with probabilities, and is already used for agent routing, guardrails, predictive spreadsheets and real-time games.

On the same data, batches and concurrency, Jev labeled 1,000 app reviews in 4.6s for $0.023 versus Gemini 3.8 Flash's 18.8s and $0.158, with 84% topic agreement and 93% bug-call agreement.