embabel / embabel-agent

Embabel Agent Framework [//]: # ([](https://sonarcloud.io/summary/new_code?id=embabel_embabel-agent)) [//]: # ([](https://sonarcloud.io/summary/new_code?id=embabel_embabel-agent))      Embabel (Em-BAY-bel) is a framework for authoring agentic flows on the JVM that seamlessly mix LLM-prompted interactions with code and domain models. Supports intelligent path finding towards goals. Written in Kotlin but offers a natural usage model from Java. From the creator of Spring.   Key Concepts Models agentic flows in terms of: • **Actions**: Steps an agent takes • **Goals**: What an agent is trying to achieve • **Conditions**: Conditions to assess before executing an action or determining that a goal has been achieved. Conditions are reassessed after each action is executed. • **Domain model**: Objects underpinning the flow and informing Actions, Goals and Conditions. • **Plan**: A sequence of actions to achieve a goal. Plans are dynamically formulated by the system, not the programmer. The system replans after the completion of each action, allowing it to adapt to new information as well as observe the effects of the previous action. This is effectively an OODA loop. > Application developers don't usually have to deal with these concepts directly, > as most conditions result from data flow defined in code, allowing the system to infer > pre and post conditions. These concepts underpin these differentiators versus other agent frameworks: • **Sophisticated planning.** Goes beyond a finite state machine or sequential execution with nesting by introducing a true planning step, using a non-LLM AI algorithm. This enables the system to perform tasks it wasn’t programmed to do by combining known steps in a novel order, as well as make decisions about parallelization and other runtime behavior. • **Superior extensibility and reuse**: Because of dynamic planning, adding more domain objects, actions, goals and conditions can extend the capability of the system, _without editing FSM definitions_ or existing code. • **Strong typing and the benefits of object orientation**: Actions, goals and conditions are informed by a domain model, which can include behavior. Everything is strongly typed and prompts and manually authored code interact cleanly. No more magic maps. Enjoy full refactoring support. Other benefits: • **Platform abstraction**: Clean separation between programming model and platform internals allows running locally while potentially offering higher QoS in production without changing application code. • **Designed for LLM mixing**: It is easy to build applications that mix LLMs, ensuring the most cost-effective yet capable solution. This enables the system to leverage the strengths of different models for different tasks. In particular, it facilitates the use of local models for point tasks. This can be important for cost and privacy. • **Built on Spring and the JVM,** making it easy to access existing enterprise functionality and capabilities. For example: • Spring can inject and manage agents, including using Spring AOP to decorate functions. • Robust persistence and transaction management solutions are available. • **Designed for testability** from the ground up. Both unit testing and agent end to end testing are easy. Flows can be authored in one of two ways: • An annotation-based model similar to Spring MVC, with types annotated with the Spring stereotype , using , and methods. • Idiomatic Kotlin DSL with and blocks. Either way, flows are backed by a domain model of objects that can have rich behavior. > We are working toward allowing natural language actions and goals to be deployed. The planning step is pluggable. The default planning approach is Goal Oriented Action Planning. GOAP is a popular AI planning algorithm used in gaming. It allows for dynamic decision-making and action selection based on the current state of the world and the goals of the agent. Goals, actions and plans are independent of GOAP. Embabel also supports Utility AI out of the box, which can run the same actions but chooses actions based on (potentially dynamic) utility scores rather than strict preconditions and postconditions. This is valuable for exploration and open-ended tasks, when we do not need to achieve a specific goal but want to maximize overall utility. The framework executes via an implementation. An agent platform supports the following modes of execution: • **Focused**, where user code requests particular functionality: User code calls a method to run a particular agent, passing in input. This is ideal for code-driven flows such as a flow invoked in response to an incoming event. • **Closed**, where user intent (or another incoming event) is classified to choose an agent. The platform tries to find a suitable agent among all the agents it knows about. Agent choice is dynamic, but only actions defined within the particular agent will run. • **Open**, where the user's intent is assessed and the platform uses _all_ its resources to try to achieve it. The platform tries to find a suitable goal among all the goals it knows about and builds a custom agent to achieve it from the start state, including relevant actions and conditions. The platform will not proceed if it is unconvinced as to the applicability of any goal. The interface provides developers a way to limit goal choice further. Open mode is the most powerful, but least deterministic. > In open mode, the platform is capable of finding novel paths that were not envisioned by developers, and even > combining functionality from multiple providers. Even in open mode, the platform will…