Summary
Cole Medin looks at practical ways to use Jev, a recently released “system one” model built to make decisions instead of generating text, to speed up AI coding workflows and cut token use. Jev takes a situation, called state, along with a set of multiple-choice questions, and answers all of them in parallel with a confidence probability for each. The video says it runs 20 to 200 times faster than an LLM and 40 to 1,000 times cheaper for these decisions.
The lead use case is security. Coding agents can be too willing to read environment variables or delete folders, even with rules in place. Jev can act as the guardrail inside a pre-tool-use hook, the lifecycle event in Claude Code and other coding agents that runs before an action. The video compares it with a Haiku-based check, which takes over a second per analysis, while Jev takes about a quarter of a second at a tiny fraction of the cost with few false positives.
Another use case is gameplay testing. A live demo shows Jev playing a game in real time, with action probabilities on screen, and catching bugs an LLM cannot find by running unit tests. The video is aimed at developers who want ideas to build into their own agent workflows.
📺 Source: Cole Medin · Published September 30, 2026
🏷️ Format: Showcase
