Jev in Practice: Browser Use and Agent Workflows

The author argues frontier models shouldn't handle YES/NO calls, and Jev should cover small decisions like agent branching, reply-or-ignore and buy/sell/hold, citing cost figures for paper classification, email sorting and browser-agent loops.

The open-source OpenComputerUse project supports background windows, lock screens and MCP, and its Recipes feature uses Cloudflare Clev or Jev models to walk the accessibility tree and execute an ordered list of actions sent by an agent.

An LLM's 30-second thinking can't keep up with candle close, so Jev returns only BUY/SELL/HOLD and a probability in 70–500ms; jev-trader, QuantDinger and AIStock are already built on it, with input priced at $0.042 per million tokens.

The author outlines a six-step loop where an LLM writes, Jev judges and code decides: packing state, asking Jev several typed questions in parallel, drafting per branch, verifying, three-tier gating and logging replayable receipts.

The post proposes using the Decisions API or Jev in Hermes to judge tool selection, model routing, failed-job priority and whether a draft needs review, with Hermes executing the chosen action; the author notes the 10x speedup applies to decision calls and real gains need testing.

The post compares Jev and the open-weight Laya, both released in September 2026, and explains which pipeline scenarios suit models that return only typed decisions with probabilities instead of generating text.

The author argues Jev's real value is in what you let it decide, and provides a full step-by-step tutorial for building your first Jev AI agent from scratch.

The team says it has been secretly running Jev in production so the agent can decide how to handle computer use operations inside the sandbox, and promises to share results and benchmarks soon.

The author uses Jev to filter papers and AI news every hour and has opened an MCP endpoint for Nexus so an agent can push the latest research summaries directly to you.

The author built a system with Jev and Groq that generates slides, charts and animations as you speak, with 70-300ms latency, and is weighing whether to open source it or charge for it.

The author says a crawler first filters out wash-traded coins, Jev makes a yes-or-no call, and Grok bots execute the trades, sharing an overnight run that turned $100 into $15,220.

Opus 5.5 sets the rules, Sonnet 5.5 writes bot.py, Jev runs an 81ms risk check before each order, and Dots watches six market feeds around the clock with no human involved.