Jev in Practice: Browser Use and Agent Workflows

A thread opener says Syftdata uses Jev for ICP, lead and enrichment scoring, and will compare per-1,000-record costs against an agent.

In ClawVille every agent uses Jev for fast decisions, so the whole town acts at once when the player does something, and players can inspect and train their agent's choices.

After two weeks of using Jev, the author finds it efficient and cheap but expects it to be embedded in harnesses and agents, where users won't notice it.

The author lays out a full Claude Code configuration tree with Opus 5.5 running the main session at low effort, per-subagent effort levels, and JEV handling millisecond micro-decisions.

Invoices, contracts, and usage records feed Jev, which decides which bills need investigation and which can reuse verified findings before handing the deep work to Haggle Bot.

A Jev paper-trading flow combines six signals into one entry score, drops launches where bundles exceed 20% or the dev holds over 5%, then sizes with Kelly and sells in tranches.

The experiment compares a Jev swarm alone, GPT6-Astra alone, and GPT6-Astra leading a Jev swarm, finding the swarm covers clues while a reasoning leader spots contradictions.

The post describes a scripted flow where Opus writes the plan, Haiku scans files in parallel, and Sonnet patches fixes, while Jev handles millisecond-level forks that need no reasoning.

Senpi launched a public feed of user prompts and replies, using Jev to classify and score each one for posting while stripping personal info.

Opus 5.5 only writes the workflow script, Haiku scans and Sonnet patches, and micro-forks that need no thinking, such as picking files or retrying versus stopping, are routed to Jev in milliseconds.

Jev acts as a routing layer that uses sharpness scores of 0.54, 0.83 and 0.89 to decide which file to edit, which agent to dispatch, and whether to retry or stop, alongside three isolated worktree agents editing code in parallel.

Corridor uses Jev to gate when security scans run and benchmarks it against decision models from OpenAI, Microsoft, Perplexity and others, with Jev 1.14 and Decider performing best at every budget.