Jev in Practice: Browser Use and Agent Workflows

Qwen writes the SQL and answers while Jev makes judgments at three points: question clarity, SQL semantics, and whether the conclusion is grounded, with Vercel eve running the loop and human approval flow.

The list includes Rizzo Flow for local Jev-style decisions and LLM2Jev for a Qwen decision model, among 5 repos meant to replace $700-a-month subscriptions.

Someone sorted the 1,142 GitHub repos built around Jev into 23 categories; the largest are evaluation and observability, agent tools, games and sims, and security guardrails, while voice chat has only 5.

The prompt has deterministic code own every quote and order, with Jev judging order-flow toxicity off the order path, and uses hftbacktest to replay ticks and check whether it beats a plain baseline.

Four agents split scanning, planning, sandboxed order execution and final sign-off, with Jev running each trade in a sealed sandbox first; one run blocked 487 bad trades.

The author uses a crawler to catch signals, Jev to make buy-or-sell calls under a second, and Grok bots to execute, turning $100 into $34,096 while he slept, with each Jev decision listed.

The author checks what Jev 1.13 actually does: it takes context and typed questions and returns probabilities, used for ticket routing, tagging messy input and score thresholds, with input at $0.042/M tokens and output free.

An open-source project lets GPT-Live talk while Jev picks faces, gestures and dances in about 120 milliseconds, all rendered in the browser.

Three changes — switching to Sonnet 5.5, trimming context with Dots and having Jev send 10 calls in parallel with an 800-token cap — cut a task from 60 seconds and $0.30 to 6 seconds and $0.03.

Instead of a conventional reranker, Jev judges each retrieved candidate against a typed question like whether the passage helps answer the query and returns a probability, dropping passages below a threshold and checking whether the evidence is sufficient.

The author chains six agents, each owning one department, into a pipeline where actions like sending email, spending money and publishing content must pass an approval layer before execution.

Pi 1.0 uses Codemode to move MCP tool orchestration into a WASM sandbox, and the post includes test code that has Jev drive a game engine through 30 decision steps.