Rod Johnson's Embabel Brings Deterministic GOAP Planning to Enterprise AI Agents on the JVM
Most AI agent frameworks hand control flow to an LLM, which makes every decision a token-burning, non-deterministic gamble. Embabel replaces that gamble with a deterministic planner that is auditable and cuts LLM costs by up to 70%, making it feasible to run agents inside regulated enterprise systems where a wrong step is unacceptable.
Embabel, a new Apache 2.0-licensed framework from Spring founder Rod Johnson, applies a game-AI planning algorithm called GOAP to enterprise AI agents. Instead of letting an LLM decide what to do next—burning tokens and introducing randomness—a deterministic planner computes the optimal sequence of Actions based on preconditions, effects, and cost. The planner itself never calls an LLM, cutting total LLM invocations by 40–60%.
All interactions are strongly typed domain objects, not stringly-typed dicts or hand-rolled JSON. Actions, Goals, and Agents are defined as plain Java or Kotlin classes with annotations, and the framework plugs directly into Spring Boot. The type system eliminates the context-bloat that inflates token usage in Python agent frameworks, saving another 20–40% in token costs.
Embabel reached 1.0.0 GA in July 2026. It positions itself as a higher-level programming model than Spring AI—where Spring AI is the parts box for calling models, Embabel is the assembly line that decides how to orchestrate them. The trade-off is a young ecosystem and a learning curve around GOAP concepts, but the payoff is auditable, cost-controlled agents suitable for finance, compliance, and multi-step enterprise automation.
GOAP shifts the trust boundary: the planner is a classical algorithm that can be stepped through with a debugger, while the LLM is confined to individual Action implementations where its output can be validated against a typed contract.
Embabel’s architecture is a direct rebuttal to the Python agent ecosystem’s reliance on LLM-driven control flow, betting instead that deterministic planning plus strong typing produces safer, cheaper enterprise agents.
The claim that the JVM is a better platform for production AI agents than Python rests on type safety and the existing Spring infrastructure, not on raw ML performance—this is an argument about software engineering, not model serving.
Because the planner is decoupled from the LLM, Embabel agents can be audited step-by-step without reconstructing prompt histories or interpreting model reasoning, which matters for SOC 2, SOX, and similar compliance regimes.