LangChain ships a Jev harness — routing and auto-mode middleware
On 17 September 2026, LangChain published Building a Harness with Jev (Sydney Runkle & Hunter Lovell). The package is langchain-typesafe.
Core surface: TypeSafeClassifier — pass state + typed questions (Noul / Choice / Score) to .invoke() and get probabilities, not chat text. State can be text, structured data, or LangChain messages.
Two experimental middleware hooks sit in the agent loop:
| Middleware | Role |
|---|---|
ModelRouterMiddleware |
Jev picks a cheap vs powerful model from your criteria before the run |
AutoModeMiddleware |
Jev screens risky tool calls (e.g. bash) before execution |
The post frames Jev as a complement to the main LLM: structured decisions inside the loop, generation left to System Two. Speed/cost multiples quoted from TypeSafe remain UnverifiedClaim here.
Primary source: langchain.com/blog/building-a-harness-with-jev.