Jev in Practice: Browser Use and Agent Workflows

The author lists all of OpenAI's DevDay announcements, describing the Decisions API as a real-time decision engine that reduces the model to picking from options, and notes Jev followed within two weeks.

A visual summary of DevDay releases including dots, GPT-6.1 Sol and Codex in the cloud, with the Decisions API letting AI choose only among given options, which the author sees as Jev being copied and its mission ending.

OpenAI's developer conference introduced the Decisions API, which picks an answer from given options instead of generating text, runs on the smallest and cheapest Luna model, returns results in about 150 ms and accepts image input; social media widely joked it copies Jev, and the post compares the two.

A full tutorial on building with Typesafe Jev from scratch, using Null/Score/Choice primitives plus confidence-based routing to make resume screening, ticket triage and content moderation demos with latency under 500 ms.

The author uses Opus 5.5 to generate trading strategies and Jev to verify each output before any money moves, and shares a 6-page paper and a Rust codebase.

The team says it has started using Jev in daily workflows at Catena Labs with good results so far, and links a long post on how it is used.

A Jev-driven trading agent completed 302 intraday round trips over 45 days with zero overnight positions, a 59.3% win rate, and grew $1,000 to $5,373.47.

The author suggests auditing sentences an agent writes just to say yes or no, and routing three checks to Jev: tool-call approval, model tier choice, and whether a task is really done, at about $0.42 per 10,000 decisions versus roughly $300 via the main model, plus a paste-ready audit prompt.

A local Chrome tool listens to speech and uses Jev to decide which line, diagram or number is worth showing and for how long, with no paid API calls; the author built it with Claude Code and is sharing the process publicly.

A shareable prompt for coding projects that gives agents a loop of Sol writing, Jev deciding, tools acting and checks verifying, so a task isn't called done just because an agent says so.

A config post that connects Claude to Jev, routes each task to the right agent and, the author claims, upgrades Claude usage beyond 99% of users.

The author suggests asking your agent how codemode and Jev can be useful, and describes extending a game built over Christmas to dump frame state so it can drive the AI with Jev via codemode.