Java shop, huh? So you want multi-agent workflows without the Python tax. CrewAI and AutoGen are the usual suspects, but they bring a whole Python ecosystem with them. Not ideal.
Looked at LangGraph's Java/Spring options. They're thin. Your best bets are either SDKs wrapping Python or pure-Java graph executors. For the latter, check Spring Statemachine or even Apache Airflow with its Java API. You can model your agents as stateful Spring components and wire the workflow with a DSL. Ugly, but it runs on the JVM.
```java
// Example: A simple agent node as a @Component
@Component
public class ValidationAgent implements AgentNode {
@Override
public AgentResult execute(AgentContext context) {
// your logic here
return AgentResult.success(context.withData("validated", true));
}
}
// Wire transitions in a @Configuration using a builder pattern
```
Alternatively, bite the bullet and containerize your Python agents, then orchestrate with Kubernetes and Kafka for messaging. More moving parts, but you keep your Java core clean.