Building Production Agents with Jev and LangGraph

LangChain published a post by Sydney Runkle and Hunter Lovell describing how its open source LangGraph framework orchestrates Jev, the decision model released by TypeSafe AI the previous week. Unlike traditional LLMs, Jev does not generate text; according to the post, it makes decisions that code can act on directly, part of what TypeSafe calls AI-powered software, in which code owns the workflow and AI handles narrow, structured decisions.

TypeSafe's benchmarks show Jev running up to 200x faster and 400x cheaper than leading LLMs on narrow decision tasks such as the routing and classification steps many agents are built around, the post states. Jev is described as a system one model, or decision model: given state and a set of questions, it returns typed answers with probabilities. The listed properties are that answers are structured, questions can run in parallel against the same state, decisions are cheap enough to make many per run, and results are designed to be stable across repeated evaluations. In an early Jev-as-a-judge experiment, the post says its scores barely moved across 100 repeated runs, less than any LLM judge tested.

LangGraph is presented as handling orchestration and reliability for such systems. A LangGraph application is built from nodes (units of work), state (information nodes read and update), and edges (which node runs next). The runtime offers durable execution through checkpointing that resumes a failed run with decisions already made, interrupts for human review, and observability through traces in LangSmith.

In a document review example for litigation, Jev answers three questions per page in one request: whether the page is responsive, whether it contains personal information, and whether it might be privileged. Answers route the page: set aside, redacted by an LLM, or sent to attorney_review with a human in the loop. The post says the same graph run with Sonnet as the judge was 5-6x slower than Jev on the classification step across trials.

The post also cites Browserbase rebuilding Stagehand's act() around Jev; in early testing, median act() latency dropped from 1.97 seconds to 0.46 seconds, about 4.3x faster, with actions below a 0.7 confidence threshold falling back to an LLM. LangGraph is described as downloaded more than 60 million times a month.