Jev in Practice: Browser Use and Agent Workflows

A demo runs multi-object detection through Jev loops with feedback, shown in video.

Decisions are packed as Choice, Score and Noul queries to Jev; those above 0.85 route to code, the rest to a human queue, with zero output tokens.

Packets route through three Jev lobes with a 0.85 gate, running 10,082 decisions at $0.42 with every packet exiting where it entered.

A leaked Jev setup splits sales outreach into brief, source, qualify and human-gate steps, with every send waiting for approval.

The author shares a one-month update on Hydra running with Jev, with a screenshot and a link to the project site.

A roundup of probability-output decision models like TypeSafe Jev, Perplexity Decider, Cloudflare Clef and OpenAI Decisions API, plus a recommended evals read.

The author wires Opus 5.5, Grok Bot, REA, Dots, Jev and Obsidian into one team with a $20 budget for 12 hours, saying Jev builds and ships while the run ends with 47 paying shops at $564/month.

Describes an /ultrareview multi-model review flow and how Jev resolves no-thinker forks in code so frontier models only see decisions that split.

The author lists marketing use cases such as AI search visibility audits, competitor ad creative gaps and CRO audits, claiming Jev Images cuts a ~$200 audit to $0.41 for outbound and lead-gen tools.

The post walks through .claude/settings.json hooks that auto-run formatters and tests at edit, submit and stop points, and suggests routing deterministic forks like tool choice and retries to Jev while big models handle reasoning.

Seven steps build a quant desk with persistent memory, researcher/backtester/reviewer subagents and a browser, with Jev making small calls in 70–500ms.

An argument that repeatedly building with Jev rewires how you model software as intelligent primitives, surfacing ideas you previously thought impossible; methodology, no concrete demo.