LangChain Integrates TypeSafe AI's Jev Model

• TypeSafe AI's Jev model claims up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks by returning typed answers and probabilities instead of generating text.
• The architecture relies on reinforcement learning for calibrated decisions (RLCD) to evaluate states and answer questions in parallel, meaning adding multiple queries barely increases response time or cost.
• LangChain's AutoModeMiddleware uses Jev to gate tool calls like bash commands, effectively extracting the closed-source safety classifiers from proprietary coding harnesses like Claude and Codex to block risky actions before execution.
• The integration supports model routing via ModelRouterMiddleware, which assesses incoming requests to dynamically select between a cheap, fast model for simple lookups and a powerful model for high-stakes architecture decisions.
• Jev supports three specific question types: Choice (picking from options), Score (rating against ordered levels), and Noul (yes-or-no probabilities), allowing software to make structured decisions without sequential LLM loops.
• The post cites Kyle Jeong from Browserbase using Jev to power browser agents for fractions of a cent, alongside Jarrod Watts building a live trading agent and Ryan Vogel doing email triage at scale.

Source: https://www.langchain.com/blog/building-a-harness-with-jev