TypeSafe AI released Jev, a new model for fast structured decisions, and LangChain explained on September 17, 2026, how developers can use it inside agent loops. Agents usually repeat a loop in which an LLM chooses an action, a tool executes it, a model checks the result, and the process continues until completion. LangChain said tool calling and structured outputs made agents easier to integrate with software, but each agent decision can still require another model call, making the loop slow and costly.

Jev is described as a System One model (fast model for structured decisions), not a traditional LLM, because it does not generate text. TypeSafe AI says Jev delivers up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks. The model evaluates a state, such as a support-ticket message, and returns typed answers with probabilities, allowing software to act on those results without a full chat LLM call for every decision.

Jev is trained with reinforcement learning for calibrated decisions, or RLCD (training for probability-aware decisions). Developers invoke the model by sending a state and questions about that state. In TypeSafe AI's support-ticket example, a customer says they have tried to connect a Stripe account for 3 days, are losing sales, and need help ASAP; Jev returns an `is_urgent` answer with `noul` of 0.999, meaning a 99.9% probability that the message is urgent.

Jev supports three question types. Choice selects from a set of options and returns a probability for each option plus an overall confidence score. Score rates an input against ordered levels such as low, medium, and high, returning a continuous score, distribution, and confidence value. Noul answers yes-or-no questions by returning the probability that a statement is true. LangChain said System One models evaluate every question in a request in parallel, so adding questions barely changes response time and mainly adds the tokens for the extra questions.

LangChain exposes Jev through `TypeSafeClassifier` in the `langchain-typesafe` package. Developers install `langchain-typesafe`, set `TYPESAFE_API_KEY`, pass state and questions to `.invoke()`, and receive classification results rather than a chat response. The state can be text, structured data, or LangChain messages, so Jev can run from a node or middleware hook using context already available to an agent.

LangChain listed model routing and Auto Mode as use cases. `ModelRouterMiddleware` lets Jev assess a user request and choose between models, such as a fast model for direct lookups, extraction, and localized changes or a more capable model for architecture and high-stakes decisions. `AutoModeMiddleware` uses Jev to check tool calls for risky decisions and block calls before a tool executes, a pattern LangChain compared with classifiers already used in coding harnesses such as claude, codex, and cursor.

LangChain also cited early projects using Jev: Kyle Jeong from Browserbase is powering browser-use agents for fractions of a cent, Jarrod Watts built a live trading agent, and Ryan Vogel is doing email triage at scale. The post said Jev is not a drop-in replacement for an LLM because it does not generate text, but can handle classification tasks often assigned to LLMs today with lower latency and cost.

TypeSafe AI released Jev, a new model for fast structured decisions, and LangChain explained on September 17, 2026, how developers can use it inside agent loops. Agents usually repeat a loop in which an LLM chooses an action, a tool executes it, a model checks the result, and the process continues until completion. LangChain said tool calling and structured outputs made agents easier to integrate with software, but each agent decision can still require another model call, making the loop slow and costly.

Jev is described as a System One model (fast model for structured decisions), not a traditional LLM, because it does not generate text. TypeSafe AI says Jev delivers up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks. The model evaluates a state, such as a support-ticket message, and returns typed answers with probabilities, allowing software to act on those results without a full chat LLM call for every decision.

Jev is trained with reinforcement learning for calibrated decisions, or RLCD (training for probability-aware decisions). Developers invoke the model by sending a state and questions about that state. In TypeSafe AI's support-ticket example, a customer says they have tried to connect a Stripe account for 3 days, are losing sales, and need help ASAP; Jev returns an `is_urgent` answer with `noul` of 0.999, meaning a 99.9% probability that the message is urgent.

Jev supports three question types. Choice selects from a set of options and returns a probability for each option plus an overall confidence score. Score rates an input against ordered levels such as low, medium, and high, returning a continuous score, distribution, and confidence value. Noul answers yes-or-no questions by returning the probability that a statement is true. LangChain said System One models evaluate every question in a request in parallel, so adding questions barely changes response time and mainly adds the tokens for the extra questions.

LangChain exposes Jev through `TypeSafeClassifier` in the `langchain-typesafe` package. Developers install `langchain-typesafe`, set `TYPESAFE_API_KEY`, pass state and questions to `.invoke()`, and receive classification results rather than a chat response. The state can be text, structured data, or LangChain messages, so Jev can run from a node or middleware hook using context already available to an agent.

LangChain listed model routing and Auto Mode as use cases. `ModelRouterMiddleware` lets Jev assess a user request and choose between models, such as a fast model for direct lookups, extraction, and localized changes or a more capable model for architecture and high-stakes decisions. `AutoModeMiddleware` uses Jev to check tool calls for risky decisions and block calls before a tool executes, a pattern LangChain compared with classifiers already used in coding harnesses such as claude, codex, and cursor.

LangChain also cited early projects using Jev: Kyle Jeong from Browserbase is powering browser-use agents for fractions of a cent, Jarrod Watts built a live trading agent, and Ryan Vogel is doing email triage at scale. The post said Jev is not a drop-in replacement for an LLM because it does not generate text, but can handle classification tasks often assigned to LLMs today with lower latency and cost.