Jev in Practice: Browser Use and Agent Workflows

After watching Jev's Dota/CS2 demo, which read 3,800+ posts in 24 seconds and produced 45k typed decisions per player, the author notes that pre-match information leaks early and asks when it can be tested on real positions.

A roundup of open-source tools such as fast-jev-compaction, typesafe-ai/skills, claude-jev, jev-engineering and jev-tree that embed typed decisions into agents, with usage data.

Jev screens events in real time and flags suspicious ones, passing only the cases that need deeper investigation to Opus 5.5, so most events never reach the expensive reasoning layer.

The author describes using a Devin-style cloud agent workflow to iterate on projects like jev-voice-control in a plain Ubuntu session with no repo connected.

TermiX offers two 1 USDC agent services: Jev FC Buddy, where an AI plays as your NES 2P teammate, and Holo Card Studio, which generates 3D holographic cards, with 1000 TP for completing and sharing.

The author built juna, a profile for pidotdev that uses Jev and Exa web search and fetch, and claims it is more token-efficient than pi.

Every describes Jev as a fast, cheap model that returns a probability for a question about text or structured data, and @hammer_mt used it to build Jevgram, which scores whether text sounds like AI.

Across 44 benchmarks and 7,193 responses, JEV detected failures such as hallucinations and prompt injection with a median AUROC of 0.886, at $0.30 per check versus $18.96 for an LLM judge.

The author ran Jev as a decision-level judge over 44 benchmarks and 7,193 responses, reaching a zero-shot median AUROC of 0.886 and using confidence cascades to raise accuracy to 0.933.

The author breaks coding guidelines into small rules Jev can judge in a single fast call, validates them with a synthetic eval, and uses Jev's confidence scores to send medium-confidence violations to the agent for review, roughly doubling recall with only a small rise in false positives.

The author gave Opus 5.5 a single prompt to make a 20-second motion graphics launch video for the open-source project jev-browser-skill, and both the visuals and the background music came out surprisingly well.

Tired of low-quality Jev content, the author built a product research tool with Jev as the judgment layer, connected to Shopify UCP and the Meta ads library, available by commenting JEVSLOP.