Jev in Practice: Browser Use and Agent Workflows

An AI safety application of Jev: rapid search over massive amounts of agent data in cases where LLMs would be impractically expensive.

Across 10,000 verifier decisions, Jev was 6.2x cheaper and 34x faster than GPT-5 Mini, with 88% agreement with human labels.

A roundup of five open-source projects using Jev for binary decisions: debate rhetoric classification, ALLOW/ASK/DENY gating for Claude Code commands, pre-trade approval, spoken English correction, and a 0.6B NanoJev clone that beat the original on a Doom test.

A SpaceXAI engineer describes replacing hiring with a graph of multiple GrokBots, and the post links to an article on how to become a Jev engineer.

At Agent Conf, Nick Miller shows a team of 20+ bots each with its own computer and points readers to an article on becoming a Jev engineer.

A Python adapter feeds enemies, health and ammo to the models, which only choose move left, move right or fire; across eight seeds Jev averages 5.63 kills, Qwen 3.5 4B 3.63, and Laya and ModernCE 1.25 each.

Two setup paths are offered: add a Jev decision layer to Claude Code, or use the Keel harness, so Jev makes narrow decisions and Opus runs at full power only when it is worth it.

Jev scores pools in 190ms, but each pool is scored from scratch with no memory, so a 2,000-neuron bee brain was added to remember which combinations ended badly.

The official skill reorders Jev calls to docs, behaviour, judgments, then a single request, and merging 13 questions into one call cut the bill 12.2x.

Layers such as Gatekeeper, Router and Worker let Jev only route, score, block and approve, with claims of 19,000% faster speed and 99.7% lower cost.

The author argues that request-level routers like Cursor and OpenRouter+Jev lack context signals and are hurt by prompt caching, so routing should happen at the harness layer per task with multi-agent collaboration.

A prompt template where Open SWE makes scoped changes and Jev chooses the next step among inspect, revise, verify and prepare_pr, producing a draft PR with test evidence.