Spring Founder Returns with Embabel: A Production‑Grade Java Agent Orchestration Framework

Java veterans' "Agent anxiety"

For the past two years AI agents have surged, but Java developers have felt stuck: the mainstream options (LangChain, CrewAI, LangGraph) live in Python, forcing Java teams either to adopt multi‑language stacks with doubled operational cost or to hand‑craft orchestration logic on top of Spring AI.

Rod Johnson sees the gap

At JDConf 2026 Rod Johnson announced Embabel (pronounced Em‑BAY‑bel), an Apache‑2.0 open‑source project released on July 20. He argues that the real problem is the gulf between AI's promises and enterprise realities, where a "90% effective, 10% ineffective" model is unusable for finance, order processing, or compliance.

Embabel’s goal is to embed nondeterministic LLMs into a deterministic engineering system.

Core positioning: Servlet vs Spring MVC analogy

While Spring AI operates at the Servlet API level, Embabel sits at the Spring MVC level, providing an orchestration layer that lets developers declare goals and typed actions, letting the framework compute execution order.

Key concepts

GOAP (Goal‑Oriented Action Planning) : a game‑AI planning algorithm where a set of actions with preconditions and effects are combined to achieve a declared goal.

OODA loop (Observe‑Orient‑Decide‑Act) : after each action the framework re‑observes state and replans, handling tool failures or new information without crashing.

The planner does not rely on an LLM; classic AI planning determines the next step, making the process deterministic, debuggable, and near‑zero cost, while LLMs handle only fuzzy transformations.

Strong‑type backbone

Embabel is written in Kotlin but feels natural in Java. An agent application is essentially a Spring Boot app; agents are Spring beans, inheriting dependency injection, transactions, security, and persistence for free.

Its programming model revolves around three pillars:

Domain model : Java record s define the data flowing through the agent. The framework infers pre‑ and post‑conditions from action input/output types, eliminating untyped maps and manual state‑machine maintenance.

Actions & goals : annotations @Action and @AchievesGoal let developers declare "what can be done" and "what must be achieved" instead of hard‑coding step sequences.

Model routing : built on Spring AI, Embabel supports OpenAI, Anthropic, Gemini, DeepSeek, Ollama, etc. Model selection can be fine‑grained per action, allowing a mix of strong and cheap models via role aliases like "best" or "cheapest".

These strong‑type features bring testability, observability, transaction support, and explainable execution paths, fulfilling Rod Johnson’s demand that every agent step be observable, reproducible, and auditable.

Comparison with Python ecosystems

Python frameworks such as LangGraph adopt a "graph" approach where developers pre‑wire the entire graph; this works for static flows but fails when new branches or tool failures appear. Embabel instead uses runtime planning, discovering action sequences on the fly, and can mix GOAP planning with explicit state machines.

Compared to Akka’s Agentic Platform, which focuses on infrastructure (actor model, fault tolerance, persistence), Embabel focuses on the programming model (declared goals and actions), addressing a different layer of the problem.

Who should hop on now

Teams already using the Spring stack who want to embed AI agents with minimal marginal cost.

Financial, governmental, or enterprise services that require determinism and auditability; Embabel’s explainable planner and strong typing hit those needs.

Java developers entering AI application development who prefer to stay in Java and leverage their existing engineering practices.

Embabel is fresh (1.0 released) and its ecosystem is still growing, so early adopters may encounter limited community solutions.

Enterprise‑grade agents no longer have to be Python‑only.