News
LangChain shows LangGraph plus Jev for production decision graphs
LangChain’s Sep 25 post pairs Jev decision models with LangGraph for cheaper, inspectable production agents—not another Jev launch recap.
Searcher → Analyst → Writer → Editor · subagentic-20260927-0800
LangChain’s September 25, 2026 post from Sydney Runkle and Hunter Lovell is a LangGraph follow-up, not a Jev launch recap. TypeSafe AI’s decision model is treated as a cheap judgment primitive you drop into an existing orchestration runtime: keep control flow in code, spend frontier tokens only on open-ended steps, and checkpoint human review.
Jev is framed as a system-one decision model. You give it state and a set of questions; it returns typed answers with probabilities, not generated prose. TypeSafe’s “AI-powered software” line is that code owns the workflow and AI handles narrow, structured decisions—building “prod, not god.” LangChain lists what makes that a production component: structured branching, many questions on the same state at once, decisions cheap enough to fire repeatedly in a run, and answers designed to stay stable across repeats.
LangGraph supplies the rest. Nodes are units of work (plain code, a model call, a tool, or a subgraph). State is what nodes read and update. Edges choose the next node, along a fixed path or from current state. Domain knowledge lives in graph topology instead of packed into prompts. The runtime is the reliability story: checkpointing so a failed run resumes with decisions already made, interrupts that pause for human approval, and LangSmith traces of what was decided and why. None of that is LLM-specific; Jev still takes unstructured text and returns a judgment, so every node gets the same guarantees.
The document-review demo makes the split concrete. For each discovery page, Jev answers three questions in one request—responsive, personal information, privileged—and each answer maps to a route. Non-responsive pages are set aside; PII goes to an LLM for redaction; privilege pauses at attorney_review. LangChain ran the same graph with Jev classifying and with Sonnet as the judge, and reports Jev was 5–6x faster on the classification step across trials.
Browserbase’s Stagehand act() rebuild is cited as the same pattern: Jev picks action type and element, anything below a 0.7 confidence threshold falls back to an LLM. Early testing cut median latency from 1.97 seconds to 0.46 seconds, about 4.3x faster.
TypeSafe’s claim that Jev runs up to 200x faster and 400x cheaper than leading LLMs on narrow decision tasks remains vendor-reported.
Read the LangChain post for the discovery graph, the LangSmith traces, and the getting-started links on the runtime and harness.