Open-source Java 21 framework for orchestrating teams of AI agents. Typed output, parallel workflows, hierarchical delegation, MapReduce, review gates, and full observability. Built on LangChain4j.
Joined on
Meet the team
Our story
We built AgentEnsemble because Java teams shouldn't need Python to orchestrate AI agents. The multi-agent ecosystem was Python-first, leaving enterprise teams with bad options: bolt on a sidecar or build from scratch.
AgentEnsemble is a Java 21 framework that brings multi-agent orchestration natively to the JVM. Define agents with roles and goals, wire tasks with typed dependencies, and let the framework handle execution -- sequential, parallel, hierarchical, or MapReduce. Built on LangChain4j, it works with any LLM provider and integrates with your existing build, monitoring, and deployment stack.
Our stack
Java 21 with virtual threads for concurrent agent execution. LangChain4j for LLM provider abstraction -- OpenAI, Anthropic, Google, Ollama, Azure, and Bedrock all work out of the box.
Gradle with Kotlin DSL for builds. JUnit 5 and AssertJ for testing. SLF4J for structured logging. Micrometer for metrics export to Prometheus, Datadog, or CloudWatch.
The docs site runs on Astro with Starlight. The live execution dashboard is React with Vite. E2E tests use Playwright.
Ships as standard Maven Central artifacts. No native dependencies, no Python runtime. Just add the JAR.