Routing Agents with Jev and Laya: Adding System One Decisions to Quarkus LangChain4j

Kevin Dubois rewires a Quarkus LangChain4j car-rental advisor so agent routing and reply checking use Jev-style System One decisions instead of chat completions. It is part how-to and part effort report, with code, on Java 25, Quarkus 3.39.x and LangChain4j 1.14.x.

A @PlannerAgent replaces the built-in LLM supervisor. The custom JevRoutingPlanner reads the customer message, asks one Choice question over criteria for each specialist, writes the route into state, and returns the action that invokes the matching sub-agent.

With an OutputGuardrail that asks one Noul question: does this reply actually answer the user? Noul returns a yes-probability between 0 and 1, so a low value means rejecting the reply. The post notes the original app had no check on whether the right agent was called or replies were good.

Choice picks one from a list and returns the choice with probabilities and confidence; Score rates on a scale and returns the score, legend, probabilities and confidence; Noul answers is-this-true with a yes-probability between 0 and 1.

No. TypeSafe paused signups, so the author did not have an API key; the post's Jev runs are a deterministic stub, while the Laya runs are real. Because the two share a schema and interface, swapping them is a one-property config change.

About 285 lines of actual decision logic. The post says the hooks were already there — a Planner to pick the agent and an OutputGuardrail to reject bad replies — and everything else is Quarkus orchestration, retries and the REST layer.

A Quarkus service on Java 25, Quarkus 3.39.x and LangChain4j 1.14.x, based on the Quarkus LangChain4j workshop app. The sub-agents it can route to are ReservationAgent, WeatherAgent, CostAgent and GeneralAgent.