Jev in Practice: Browser Use and Agent Workflows

The author shares screenshots of the Jev model in real use, showing how it performs in practice.

After Jev performed poorly in earlier internal tests for semantic model routing, the author reworked the implementation, such as adding the previous round's decision to the decision chain, and found results much better; the latest Magpie now supports Jev for semantic model routing.

The author has integrated Jev into the Mnemos cognitive engine, is tuning its granular mechanics through recursive testing, and calls the Jev plus Opus 5.5 combination very strong.

The Green Eggs agent can generate Python scripts that control a subagent through RPC, and Jev-powered comparators are planned for automated research orchestration.

A 14-page PDF compiled from TypeSafe and Anthropic docs lays out an architecture where Opus 5.5 thinks and Jev decides, with Choice, Score and Noul evaluated in parallel and branching on probability to skip most generation calls.

The author routed 1,948 decisions from 07:00 to 23:00 through Jev in about 0.1 seconds each, with Opus 5.5 doing the work and irreversible actions held for approval, claiming 3 hours 40 minutes saved.

Jev Router scores each conversation for difficulty and benefit to decide when to switch models or add reasoning, solving 237 of 423 tasks on four agent benchmarks versus 130 for OpenRouter's Auto Router.

A directory of more than 50 deterministic control systems covering model routing, context compaction, browser sandboxes, guardrails and human escalation gates, all organized around LLMs proposing, Jev deciding and code enforcing.

A Japanese-language AI news roundup whose items include OpenRouter introducing the Jev-based Jev Router to pick models by task difficulty, and TypeSafe AI reportedly in talks for a funding round at a ten-billion-dollar valuation.

Ollaya runs a lightweight decision model on a Mac GPU or CPU with one command, returns probabilities instead of text, and offers MCP support for Cursor and Claude Code.

Autorubric 1.6.0 now supports all Jev-shaped models in addition to LLMs, and Jev users need to install autorubric[typesafe].

A video has Jev and Eikos, an open-source project a Brazilian developer built in three days, play Tetris side by side, with Eikos winning 5-0.