Comparisons· Last updated 2026-09-21

LangChain harness patterns with Jev

On 17 September 2026 LangChain published Building a Harness with Jev. It is a LangChain document, not a TypeSafe one. It describes where a System One model sits in an agent loop: LLMs still generate and call tools; Jev makes the cheap structured decisions.

They quote TypeSafe’s “up to 200x faster / 400x cheaper” class of claims. Those remain vendor-reported. See Jev vs LLM.

Package surface (as of that post)

from langchain_typesafe import Noul, TypeSafeClassifier

classifier = TypeSafeClassifier()
response = classifier.invoke({
    "state": "The deploy failed twice and customers are seeing 500s. Can someone look now?",
    "questions": {
        "urgent": Noul(instructions="Does this need attention right now?"),
    },
})
urgency = response.nouls["urgent"].noul

State may be text, structured data, or LangChain messages.

Model routing

ModelRouterMiddleware plus ModelChoice entries: Jev picks fast vs powerful (or whatever you name) from the latest user message.

Auto mode

AutoModeMiddleware uses Jev to inspect tool calls (the post’s example lists bash) and block risky ones before execution. LangChain frames this as opening a pattern that closed harnesses already use internally.

Install langchain-typesafe, set TYPESAFE_API_KEY, and follow current LangChain docs if the experimental middleware paths change.

jev.pro is not a LangChain affiliate either.