Jev in Practice: Browser Use and Agent Workflows

The author argues against building one all-purpose agent, instead giving each agent a single job with small focused context, running work in parallel, repairing failures without restarts, and merging only verified results, which they say is faster, cheaper and more reliable with Claude Code.

The author says trading with Jev still shows no edge while testing 500+ strategies distilled from 2.1 million evaluated strategies, 19.0 million Monte Carlo simulations and 300+ analyzed markets.

The author shares practical notes on plugging the Typesafe model into coding agent workflows.

Diogo Almeida draws the line between assistance AI and automation AI, and explains how Jev replaces human approval with calibrated decisions.

A Claude Code mod uses Jev to automatically classify the subagent model, main model and effort level for each request, installable with one command.

The author used Typesafe to visualize a research and capital map of robotics companies, with the code open-sourced on GitHub.

In a chief-of-staff style orchestrator, Jev replaces the main LLM's context reading and autoregressive decision-making to make subagent routing decisions more efficiently.

Someone built voice-controlled computer use for Mac with Jev that executes commands faster than they can be spoken, opening apps before the sentence is finished.

The author posts a video demonstrating a new use case they found for Jev.

Rippling tested Jev on anonymized real-world data in GrowthOS: it reached 90% classification accuracy versus a 50% LLM baseline and ran nearly 8x faster, but performed poorly as a control mechanism in its text-to-SQL agent loop.

The author built a Claude Code mod based on TypeSafe's Skill Suggestion example that hands skill, effort and tool selection to decision models like Jev, splitting the harness into independent decision services.

Astra plans goals and distills skills while Jev outputs movement and combat actions in real time, for a total run cost of about $0.97.