A Type-Safe Agent Framework From the Creator of Spring

What it does

Embabel is a Kotlin framework for building agentic flows on the JVM that interleave LLM calls with regular code and domain objects. It treats actions, goals, and conditions as strongly typed first-class citizens, letting Spring inject and manage agents like any other bean. The system formulates plans dynamically and replans after each action, so it can adapt to new information rather than following a predetermined script.

The interesting bit

Instead of chaining LLM calls in code or wrestling with finite state machines, Embabel uses a non-LLM planning algorithm—by default GOAP, the same technique used in game AI—to figure out how to reach a goal. Adding new capabilities can extend system behavior without editing existing flow definitions, because the planner invents new action sequences at runtime.

Key highlights

• Strongly typed domain model with full refactoring support; no “magic maps” passing unstructured data between steps.
• Pluggable planning layer ships with both GOAP (goal-oriented) and Utility AI (score-based) strategies out of the box.
• Three execution modes: Focused (call a specific agent), Closed (classify intent and pick a known agent), and Open (assemble a custom agent from all available goals and actions).
• Built on Spring, so agents participate in dependency injection, AOP, transactions, and existing enterprise infrastructure.
• Supports mixing LLMs—local models for cheap point tasks, larger models for heavy lifting—to balance cost and capability.

Caveats

• Natural language actions and goals are still on the roadmap, not yet available.
• Open mode explicitly trades determinism for power, and the framework is currently at a 0.1.2-SNAPSHOT maturity level.
• Federation and “evolving” multi-goal modes are listed as future work, not current features.

Verdict

Worth evaluating if you are building JVM-based agent systems and want type safety plus dynamic planning without leaving the Spring ecosystem. Skip it if you just need a simple linear chain of LLM calls.