TypeSafe AI Launches Jev to Speed Up AI Agent Decisions

TypeSafe AI has launched Jev, a specialized AI model designed to streamline the decision-making loops of autonomous agents. Unlike traditional large language models, Jev is a System One model that does not generate text. Instead, it evaluates a given state and returns structured, typed answers and probabilities. By focusing purely on classification rather than text generation, Jev achieves up to 200 times faster inference and 400 times lower costs than comparable language models on classification tasks.

The model is trained using reinforcement learning for calibrated decisions, or RLCD. Developers can query Jev by sending it a state context alongside specific questions, which the model evaluates in parallel to keep latency low. Jev supports three distinct question types: Choice, which selects from options and returns confidence scores; Score, which rates inputs against ordered levels to output a continuous score; and Noul, which answers yes-or-no queries with a truth probability.

To help developers implement this technology, LangChain has introduced integration support through the langchain-typesafe package. Using the TypeSafeClassifier, developers can invoke Jev to handle routing and guardrails. For instance, Jev can act as a router to select the most cost-effective model, such as deciding when to pass a complex task to openai:gpt-5.6-luna. It can also serve as safety middleware, using AutoModeMiddleware to analyze tool calls like bash commands and block risky actions before they execute.

Early adopters are already utilizing Jev for various high-speed applications. Kyle Jeong from Browserbase is using the model to power browser-use agents for fractions of a cent, while Jarrod Watts has built a live trading agent. Additionally, Ryan Vogel is employing Jev to manage large-scale email triage. By offloading structured classification from expensive generative models, Jev provides a fast, secure, and highly economical alternative for agent architectures.