Jev in Practice: Browser Use and Agent Workflows

ReproForge follows an arXiv study and swaps in Jev at bounded decision points in agent workflows to test whether faster local decisions actually improve end-to-end agent performance.

The author built JevChess, a free browser-based 3D chess game with three difficulty levels against TypeSafe's Jev decision model, playable on phones, with source on GitHub.

The author ran Mac Arena with Qwen, Astra and Opus operating the computer, finding small models still far behind GPT-6 Astra, with Jev only speeding up and improving the experience.

Jev reads every note before Opus writes the important parts, which the author says cuts weekly chat time by 80% and keeps the vault working while offline.

On TermiX's agent.family, Jev FC Buddy lets a player take 1P while the AI plays 2P in classic NES co-op games, alongside a 1 USDC service that makes custom 3D hologram cards.

Jev picks the next action, code catches repeated work, and Opus 5.5 handles tasks needing deeper investigation, routed through six steps from prepare to measure.

A developer lists the tech stack behind the Regents web apps, including Jev to provide sybil resistance for agents.

After opening sign-ups outside business hours, the author added screening that combines Exa search and Jev to filter out abusive registrations from people without real company emails, and says it works well.

youdotcom partnered with n8n to build a weekday AI news brief using Jev, which makes the per-article filtering judgment at a cost of $0.00004 each.

MIT students list 9 use cases pairing Jev for quick decisions with Opus 5.5 for deeper reasoning, covering request routing, email sorting, and gating tool calls.

The author split the work across Haiku reading contracts and invoices, Jev judging each line as pass, flag or skip, and Opus writing code, reconciling 41 contracts with 3 years of invoices in a system sold for $115,000.

On day 13 of a paper run, Jev acts as a sub-second yes/no gate while CHIEF picks PEPPER with a 0.84 worth score and 0.79 confidence, but risk checks fail so the desk only watches without buying.