Jev in Practice: Browser Use and Agent Workflows

The author combines Opus's deep reasoning with Jev's fast, cheap decision-making into a trading bot architecture they call the strongest setup to date.

The author places Jev in a System 1/2 decision framework, discussing when to answer fast versus switch to deeper reasoning, and maps predicting task success to the RL Q-function.

A project built in 90 minutes uses Jev to label calendar and Notion events as no funding, dubious or spend, then has GPT Luna allocate budgets by importance and remaining funds.

A roundup of 10 Jev projects covering desktop agents, Monad trading experiments and a model playing Mario from emulator state, where Jev picks actions and surrounding code handles execution.

A developer open-sourced chip, a minimal coding agent where a decision model like Jev picks the tool and Claude fills in the arguments, looping until the decider stops; the demo audio was generated with it too.

Using Tetlock's historical forecasting problem set, the author had Jev and D1 assign probabilities to each outcome and compared them with actual results; D1 won 9 of 12 categories.

Cloudflare trained its own Clef and Clef-flash decision models, compatible with Jev's System One API, with lower latency but weaker results than Jev on knowledge-heavy tests.

Anthropic made Claude Code customizable, and the author suggests adding a Jev-style System One review layer to scan for architecture violations, bugs and spec drift.

The Jev agent ran 276 trades with a 62.3% win rate, cut every loss at -1.8R, and turned $1,000 into $9,227.71, with the losing trades charted in a separate histogram.

Cloudflare open-sourced its decision model Clef-flash, which it says is 13x faster than Jev but more than twice the price at similar accuracy, though one developer's test questions the speed claim.

Five strategies fed into the Jev decision layer: mean reversion was cut on day 4 after losing $148 and the news strategy on day 7, leaving 3 that ran 88 trades and grew $1,000 to $2,957.08.

Cloudflare released Clef, a 27B/9B typed decision model that outputs choice and score, beats Jev on several benchmarks, and now has a call example in the TanStarter template.