Jev in Practice: Browser Use and Agent Workflows

The Typesafe account quoted a user's goal of making the experience feel impossibly magical and responded that with Jev it's possible.

Asymptote Labs open-sourced Beacon, which links sessions across more than 20 tools like Cursor, Claude Code and Codex and uses Jev to cheaply turn valuable runs into reusable skills, having already surfaced 579 sessions from 5 frameworks.

A 14-page PDF explains how to build a company brain with Jev plus Opus 5.5, and the author suggests sending it straight to your agent to read.

After Apple Spotlight broke for months, the author built a search tool with Jev that finds apps and performs actions across any app, and doubles as a universal keyboard shortcut system.

A long post surveys the /v1/systemone ecosystem, including Ollama, Perplexity and AWS integrations, its choice/noul/score primitives, and a finding that adversarial injection can flip 60% of outputs.

The author handed his brokerage account to three models: Haiku 5.5 writes news cards, Jev outputs only buy/sell/skip with a probability (no action below 0.80), and Opus 5.5 sets position size and places orders, making $36,015 in two days.

The author outlines a Jev workflow: Alpaca for stock and option prices, QuantLib for a probability model, the Polymarket SDK for order books, and Jev to read settlement rules and sort candidates into INVESTIGATE, WATCH or SKIP.

Each lead found by the scouts gets a verdict and confidence score from Jev before Hydra decides whether to dig deeper, and the verdicts are graded against later outcomes so the swarm learns which calls to trust.

The author says Jev lets the agent stay with you, see your screen and write on it via Haiku, while Opus 5.5 handles reasoning, forming an ambient intelligence.

A broker CSV is fed to Jev to tag each trade, code drops 2,031 of them, and the remaining 109 go to Opus 5.5 to find patterns, which locates three losing habits written up as new trading rules.

Opus 5.5 runs the main session, Sonnet 5 handles subagent work, Haiku grades whether the goal is met, and JEV routes micro-forks like tool choice and retries in milliseconds, with a pasteable config prompt.

Sets Fable 5.1 as an advisor behind every Sonnet 5.5 subagent, stepping in only at three key forks: rewriting multiple files, repeated failures, and reporting completion.