Building a Harness with Jev
LangChain blog guide on wiring Jev into an agent harness as a dangerous-action classifier and model router.
This summary is based on how curated lists describe the project; its README was not available to review.
Overview
LangChain's walkthrough of building an agent harness with Jev as the decision layer. It argues that coding harnesses like Claude Code, Codex and Cursor already classify dangerous actions in closed parts of their code, and that a cheap, performant classifier lets the same pattern be applied to any agent.
It also shows a router that uses Jev to select a model by custom criteria while keeping probabilities in agent state, and cautions that Jev is not a drop-in replacement for an LLM. The listing notes LangChain later published its own evaluation of Jev as a judge.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this proposed agent action dangerous? | yes · no |
| Choice | Which model should handle this request? | — |
How it works
- Agent proposes an action
- Jev classifies whether it is dangerous
- Router picks a model with probabilities kept in state
- Harness acts or asks for approval
Strengths
- Opens a pattern previously locked in closed harnesses
- Keeps probabilities for later decisions
Caveats
- Guide, not a product
- Jev is not a drop-in LLM replacement
At a glance
- Type
- Article
- Status
- discussion
- Runtime
- LangChain
Use cases and integrations
How curated lists describe it
Framework guide: LangChain's walkthrough of wiring Jev into an agent harness as the decision layer, from a team that then published its own evaluation of Jev as a judge.
yibie/awesome-jev*"Coding harnesses like claude, codex, cursor have shipped some kind of way to classify dangerous actions… **Up until now, this classifier step has been locked away in the closed source parts of the harness. Now that a cheap and performant classifier model exists, we can take the same pattern and adopt it to all agents!**"*; it also provides a router example (using Jev to select a model according to custom criteria, with probabilities retained in agent state) and *"Jev isn't a drop-in replacement for an LLM"*.
wh000wh000/awesome-jev-live
Frequently asked questions
What does Jev do in the harness?
Classifies dangerous actions and routes requests to models.
Can Jev replace the LLM?
No, the post says it is not a drop-in replacement.
Where is it?
https://langchain.com/blog/building-a-harness-with-jev
Related projects
LangChain's Python framework, which includes an optional TypeSafeClassifier runnable plus model-routing and risky-tool middleware using Jev.
PluginproductionPythonLangChain.js, which includes an optional TypeSafeClassifier runnable and middleware for bounded routing and tool-call checks with Jev.
PluginproductionTypeScriptComposio agent SDK with an optional TypeSafe provider that uses Jev to shortlist and pick tools and closed-set arguments.
ProjectproductionTypeScriptClaude Code mod that adds $.jev to the engine interface so other plugins and hooks can ask Jev typed questions through six providers.
ServicebetaTypeScriptYielded Agent: durable Effect agent threads in TypeScript, listed for a TypeSafe decision provider for typed question sets.
ProjectbetaTypeScriptVercel's filesystem-first agent framework, whose experimental evaluate and autoModel paths default to Jev via AI Gateway.
ServicebetaTypeScript