LangChain Adds Jev, a Fast System One Model for Agent Decisions
LangChain has published a guide to Jev, a new model from TypeSafe AI that does not generate text but returns typed answers and probabilities for classification-style decisions. According to TypeSafe AI, it delivers up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks.
- Jev is a System One model from TypeSafe AI that returns typed answers and probabilities instead of generating text.
- TypeSafe AI reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks.
- Questions come in three types: Choice, Score and Noul, and many can be asked about one state in a single request.
- LangChain offers Jev through TypeSafeClassifier in the langchain-typesafe package.
- Experimental middleware covers model routing and an Auto Mode guardrail that blocks risky tool calls.
A model built for decisions, not text
Agents run in a loop: an LLM decides what to do, a tool executes, a model evaluates the results, and the cycle continues until the task is complete. LangChain notes that even with tool calling and structured outputs, which let models make structured requests and return structured results, the loop remains slow and costly because every decision requires another model call.
Jev is aimed at that problem. LangChain describes it as what the TypeSafe AI team calls a System One model, a class of AI models built to make fast, structured decisions that software can use directly. Such a model evaluates a state and returns typed answers and probabilities. According to the post, it is trained using reinforcement learning for calibrated decisions (RLCD). Application code then uses the results to steer the agent, without a full chat LLM call for each decision.
How Jev is queried
To call Jev, a developer sends a state, which is the context, along with questions about that state. In the support-ticket example from the TypeSafe docs, a customer message says a Stripe connection has failed for three days. The question asks whether the message conveys urgency. The response returns a probability of 0.999, which LangChain reads as a 99.9% probability that the message is urgent, something an application can use to prioritize the ticket.
Three question types are supported:
- Choice: pick from a set of options. It returns a probability for each option and an overall confidence score.
- Score: rate an input against ordered levels such as low, medium and high. It returns a continuous score, the underlying distribution and a confidence value.
- Noul: answer a yes-or-no question, returning the probability that a statement is true.
Several questions can be asked about the same state in one request. LangChain says the model evaluates all questions in parallel, so adding questions barely changes response time and costs only the tokens for the extra questions, which it calls cheap.
Using Jev with LangChain
The LangChain integration exposes Jev through TypeSafeClassifier. Developers install langchain-typesafe, set a TYPESAFE_API_KEY, and pass state and questions to .invoke(). They get classification results back instead of a chat response. The state can be text, structured data or LangChain messages, so Jev can be called from a node or middleware hook using context the agent already holds.
LangChain stresses that Jev is not a drop-in replacement for an LLM. It cannot generate text, but it can handle classification tasks that LLMs often perform today. The suggested pattern is to use an LLM for open-ended reasoning and generation, and Jev for fast, structured decisions along the way.
Model routing and Auto Mode
The post highlights two use cases built as experimental middleware. The first is model routing. A ModelRouterMiddleware lets Jev assess a request and pick a model against criteria the developer defines, for instance a fast, inexpensive model for direct lookups and extraction, and a more capable one for architecture and high-stakes decisions. The router selects a model from the latest user message and uses it throughout the run. The probabilities and confidence stay available in agent state.
The second is Auto Mode, a guardrail. LangChain argues that agents remain inherently untrustworthy because they can receive bad instructions, either naturally or from a motivated attacker. It says coding harnesses such as claude, codex and cursor have shipped ways to classify dangerous actions before they are taken, but that this classifier step has been locked inside closed source parts of the harness. With a cheap, performant classifier available, LangChain says the same pattern can be adopted for all agents. AutoModeMiddleware uses Jev to check tool calls for risky decisions and blocks calls before the tool executes.
What it means for agent builders
For developers, the pitch is fewer full LLM calls in the agent loop for decisions that are really classifications, such as urgency, routing and risk checks. The speed and cost figures are TypeSafe AI's own claims on classification tasks, and the post does not present independent benchmarks.
LangChain also points to early projects: 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. LangChain says the model drew an outsized response compared with the many models that launch each week.
LangChain: What Is Jev? A Guide to TypeSafe AI’s System One Model
