Jev in Practice: Browser Use and Agent Workflows

An MIT-licensed configuration with 68 subagents, 286 skills and 94 commands that runs a single Claude Code assistant like a full engineering team, with Jev mentioned for routing decisions.

The author looks back on a month of making 3D with Astra, demos with Jev and videos with Opus 5.5, and finds nothing left but a pile of folders and a few hundred dollars in bills.

Five Hermes plugins trending today, among them typesafe-skill-router, which uses Jev to decide which skill applies before the model speaks and injects nothing when none matches.

The author ran Jev 1.13.0 against a real inbox: typed questions in, decisions out, anything below a 0.85 confidence threshold escalated to a human, at about $0.42 per 10,000 decisions.

Jev Brain CLI can read and patch code, run tests, and open PRs on its own, distributes skills as SKILL.md packs, and routes every tool call through Agent Warden for approval.

The JevBench creator argues against picking System One models by vibes and explains how rotating metrics, public plus private question sets, and benchmaxxing detection keep scores from being gamed.

The JevBench creator rejects choosing System One models by vibe and trust, and describes how version changes, public and private test pools, and benchmaxxing detection prevent score gaming.

The 0.8B, 2B, and 4B models average 79.38, 84.68, and 90.02 across seven decision benchmarks, with the 4B reaching 88.74 to pass Jev at lower latency and returning structured JSON with calibrated probabilities in one forward pass.

The author bought a $1 Jev-powered FC buddy seat on termix_ai, where an agent plays as player 2 for half an hour and delivers watchable live inputs rather than a document.

Opus 5.5 handles research and signal generation while Jev scores each order before it is sent, checking signal agreement, market regime and existing risk, and blocks anything below the confidence threshold with a logged reason.

OpenHuman says it is the first agent harness to use Typesafe Jev, which selects tool calls, automates desktop and browser control and handles orchestration, with a demo video included.

Unsloth added a Decision API that runs Laya decision models locally on 4GB of RAM and exposes a Jev-compatible interface, so existing Jev projects can switch the endpoint with little code change.