Embabel Agent Framework

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 @Agent, using @Goal, @Condition and @Action methods.
Idiomatic Kotlin DSL with agent { and action { 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. Future planning options include:

Plans created by a reasoning model such as OpenAI o1 or DeepSeek R1.

The framework executes via an AgentPlatform 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 GoalChoiceApprover 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 only perform individual steps that have been specified. (Of course, steps may themselves be LLM transforms, in which case the prompts are controlled by user code but the results are still non-deterministic.)

Possible future modes:

Evolving mode: Where the platform can work with multiple goals in the same process and modify a running process to add further goals and agents. For example, an action can realize that it has become important to achieve additional goals.

Embabel agent systems will also support federation, both with other Embabel systems (allowing planning to incorporate remote actions and goals) and third party agent frameworks.

Quick Start

Get an agent running in under 5 minutes.

Create your own agent repo from our Java or Kotlin GitHub template by clicking the "Use this template" button.

You can also create your own Embabel agent project locally with our quick start tool:

uvx --from git+https://github.com/embabel/project-creator.git project-creator

Choose Java or Kotlin and specify your project name and package name and you'll have an agent running in under a minute, if you already have an OPENAI_API_KEY and have Maven installed.

📚 For examples and tutorials, see the Embabel Agent Examples Repository

🚗 For a realistic example application, see the Tripper travel planner agent

Why Is Embabel Needed?

TL;DR Because the evolution of agent frameworks is early and there's a lot of room for improvement; because an agent framework on the JVM will deliver great business value.

Why do we need an agent framework at all? We can write code without higher level abstractions, directly invoking LLMs and controlling flow directly in code. However, a higher level agent framework offers compelling benefits. For example:
- Breaking up LLM interactions, making them simpler and more focused. This maximizes reuse and minimizes cost and errors. It often allows us to use cheaper models for point interactions.
- Facilitating both unit and integration testing, which remain as important with agentic systems as with any other software systems.
- Increasing composability where subflows and individual actions can be reused
- Making applications more manageable and robust, enabling a workflow manager to control their execution and retry operations while maintaining previous state
- Enhancing safety through the ability to apply guardrails in many places
Why do we need an agent framework for the JVM when solutions exist in Python?: While this space is presently better developed in Python (or even TypeScript), it's early and there's plenty of room for novel and potentially superior approaches. The key adjacency is often not the LLM--which is a simple HTTP call away--but existing code and infrastructure assets that are more likely on the JVM than in Python.
Why not use just Spring AI? Spring AI is great. We build on it, and embrace the Spring component model. However, we believe that most applications should work with higher level APIs. An analogy: Spring AI exists at the level of the Servlet API, while Embabel is more like Spring MVC. Complex requirements are much easier to express and test in Embabel than with direct use of Spring AI.
Why not attempt to contribute this project to Spring? This project requires different governance from Spring, where most projects exist in stable environments and dependability and stability outweighs rapid innovation. Second, the concepts are not JVM-specific. We hope that Embabel will become the leading agent framework across platforms. While the Spring brand is valuable in Java, it is not in TypeScript or Python.

Show Me The Code

In Kotlin or Java, agent implementation code is intuitive and easy to test.

@Agent(description = "Find news based on a person's star sign")
class StarNewsFinder(
    // Services such as Horoscope are injected using Spring
    private val horoscopeService: HoroscopeService,
    private val storyCount: Int = 5,
) {

    @Action
    fun extractPerson(userInput: UserInput): StarPerson =
        // All prompts are typesafe
        PromptRunner().createObject("Create a person from this user input, extracting their name and star sign: $userInput")

    @Action
    fun retrieveHoroscope(starPerson: StarPerson) =
        Horoscope(horoscopeService.dailyHoroscope(starPerson.sign))

    // toolGroups specifies tools that are required for this action to run
    @Action(toolGroups = [ToolGroup.WEB])
    fun findNewsStories(person: StarPerson, horoscope: Horoscope): RelevantNewsStories =
        PromptRunner().createObject(
            """
            ${person.name} is an astrology believer with the sign ${person.sign}.
            Their horoscope for today is:
                <horoscope>${horoscope.summary}</horoscope>
            Given this, use web tools and generate search queries
            to find $storyCount relevant news stories summarize them in a few sentences.
            Include the URL for each story.
            Do not look for another horoscope reading or return results directly about astrology;
            find stories relevant to the reading above.

            For example:
            - If the horoscope says that they may
            want to work on relationships, you could find news stories about
            novel gifts
            - If the horoscope says that they may want to work on their career,
            find news stories about training courses.
        """.trimIndent()
        )

    // The @AchievesGoal annotation indicates that completing this action
    // achieves the given goal, so the agent run will be complete
    @AchievesGoal(
        description = "Write an amusing writeup for the target person based on their horoscope and current news stories",
    )
    @Action
    fun writeup(
        person: StarPerson,
        relevantNewsStories: RelevantNewsStories,
        horoscope: Horoscope,
    ): Writeup =
        // Customize LLM call
        PromptRunner().withTemperature(1.2).createObject(
            """
            Take the following news stories and write up something
            amusing for the target person.

            Begin by summarizing their horoscope in a concise, amusing way, then
            talk about the news. End with a surprising signoff.

            ${person.name} is an astrology believer with the sign ${person.sign}.
            Their horoscope for today is:
                <horoscope>${horoscope.summary}</horoscope>
            Relevant news stories are:
            ${relevantNewsStories.items.joinToString("\n") { "- ${it.url}: ${it.summary}" }}

            Format it as Markdown with links.
        """.trimIndent()
        )

}

The following domain classes ensure type safety:

data class RelevantNewsStories(
    val items: List<NewsStory>
)

data class NewsStory(
    val url: String,

    val summary: String,
)

data class Subject(
    val name: String,
    val sign: String,
)

data class Horoscope(
    val summary: String,
)

data class FunnyWriteup(
    override val text: String,
) : HasContent

Dog Food Policy

We believe that all aspects of software development can and should be greatly accelerated through the use of AI agents. The ultimate decision makers remain human, but they can and should be greatly augmented.

This project practices extreme dogfooding.

Our key principles:

We will use AI agents to help every aspect of the project: coding, documentation, producing marketing copy etc. Any human performing a task should ask why it cannot be automated, and strive toward maximum automation.
Developers retain ultimate control. Developers are responsible for guiding agents toward the solution and iterating as necessary. A developer who commits or merges an agent contribution is responsible for ensuring that it meets the project coding standards, which are independent of the use of agents. For example, code must be human readable.
We will use only open source agents built on the Embabel platform, and contribute any improvements. While commercial coding agents may be more advanced, we believe that our platform is the best general solution for automation and by dogfooding we will improve it fastest. By open sourcing agents used on our open source projects, we will maximize benefit to the community.
We will prioritize agents that help accelerate our progress. Per the flight safety advice to fit your own mask before helping others, we will prioritize agents that help us accelerate our own progress. This will not only produce useful examples, but increase overall project velocity.

Developers must carefully read all code they commit and improve generated code if possible.

Getting Started

Get the bits
Set up your environment
Run the application

Getting the bits

Choose one of the following:

Clone the repository via git clone https://github.com/embabel/embabel-agent
Create a new Spring Boot project and add the necessary dependencies (see "Using Embabel Agent Framework in Your Project" below)

Environment variables

Environment variables are consistent with common usage, rather than Spring AI. For example, we prefer OPENAI_API_KEY to SPRING_AI_OPENAI_API_KEY.

Required:

OPENAI_API_KEY: For the OpenAI API

Optional:

ANTHROPIC_API_KEY: For the Anthropic API. Necessary for the coding agent.

We strongly recommend providing both an OpenAI and Anthropic key, as some examples require both. And it's important to try to find the best LLM for a given task, rather than automatically choose a familiar provider.

Services

You will need a Docker Desktop version >4.43.2. Be sure to activate the following MCP tools from the catalog:

Brave Search
Fetch
Puppeteer
Wikipedia

You can also set up your own MCP tools using Spring AI conventions. See the application-docker-desktop.yml file for an example.

If you're running Ollama locally, set the ollama profile and Embabel will automatically connect to your Ollama endpoint and make all models available.

Running

Create your own agent project with

uvx --from git+https://github.com/embabel/project-creator.git project-creator

Example Agents

📚 For examples and tutorials, see the Embabel Agent Examples Repository

# Clone and run examples
git clone https://github.com/embabel/embabel-agent-examples
cd embabel-agent-examples/scripts/kotlin
./shell.sh

Shell Commands

Type help to see available commands. Example:

execute "Lynda is a Scorpio, find news for her" -p -r

Options:

-p logs prompts
-r logs LLM responses

Use chat for interactive conversations with agents.

Tip: Spring Shell supports history - type !! to repeat the last command

Type help to see the available commands.

An example:

execute "Lynda is a Scorpio, find news for her" -p -r

This will look for an agent, choose the star finder agent and run the flow. -p will log prompts -r will log LLM responses. Omit these for less verbose logging.

Use the chat command to enter an interactive chat with the agent. It will retain conversation history, and attempt to run the most appropriate agent for each command.

Spring Shell supports history. Type !! to repeat the last command. This will survive restarts, so is handy when iterating on an agent.

Further examples

Example commands within the shell:

# Perplexity style deep research
# Requires both OpenAI and Anthropic keys and Docker Desktop with the MCP extension (or your own web tools)
execute "research the recent australian federal election. what is the position of the greens party?"

# x is a shortcut for execute
x "fact check the following: holden cars are still made in australia; the koel is a bird native only to australia; fidel castro is justin trudeau's father"

Try the coding agent (separate repo) with commands such as:


x "explain this project for a five your old"

x "take the StarNewsFinder kotlin example of the agent framework. create a parallel .java package beside its and create a java version of the same agent use the same annotations and other classes. use records for the data classes. make it modern java"

x "consider the StarNewsFinder kotlin class. This is intended as an example. Is there anything you could do to make it simpler? Include suggested API changes. Do not change code"

Bringing in additional LLMs

Local models with well-known providers

The Embabel Agent Framework supports local models from:

Ollama: Simply set the ollama profile and your local Ollama endpoint will be queries. All local models will be available.
Docker: Set the docker profile and your local Docker endpoint will be queried. All local models will be available.

Custom LLMs

You can define an LLM for any provider for which a Spring AI ChatModel is available.

Simply define Spring beans of type Llm. See the OpenAiConfiguration class as an example.

Remember:

Provide the knowledge cutoff date if you know it
Make the configuration class conditional on any required API key.

Roadmap

This project is in its early stages, but we have big plans. The milestones and issues in this repository are a good reference. Our key goals:

Become the natural way to Gen AI-enable Java applications, and especially those built on Spring.
Prove the power of the model. Demonstrate that this model is the most capable agent model. See the Dunnart milestone for details. In particular:
- Demonstrate the power of extensibility without modification, by adding goals and actions
- Demonstrate the potential to become the PaaS for natural language
- Demonstrate the potential of agent federation within the GOAP model
- Demonstrate budget-aware agents, such as "Research the following topic, spending up to 20c if you are still learning"
- Integrate with data stores and demonstrate the power of surfacing existing functionality inside an organization
Take the model to other platforms: The conceptual framework is not JVM specific. Once established, we intend to create TypeScript and Python projects.

There is a lot to do, and you are awesome. We look forward to your contribution!

Application Design

Domain objects

Applications center around domain objects. These can be instantiated by LLMs or user code, and manipulated by user code.

Use Jackson annotations to help LLMs with descriptions as well as mark fields to ignore. For example:

@JsonClassDescription("Person with astrology details")
data class StarPerson(
    override val name: String,
    @get:JsonPropertyDescription("Star sign")
    val sign: String,
) : Person

See Java Json Schema Generation - Module Jackson for documentation of the library used.

Domain objects can have behaviors that are automatically exposed to LLMs when they are in scope. Simply annotate methods with the Spring AI @Tool annotation.

When exposing @Tool methods on domain objects, be sure that the tool is safe to invoke. Even the best LLMs can get trigger-happy. For example, be careful about methods that can mutate or delete data. This is likely better modeled via an explicit call to a non-tool method on the same domain class, in a code action.

Using Embabel as an MCP server

You can use the Embabel agent platform as an MCP server from a UI like Claude Desktop. The Embabel MCP server is available over SSE.

Configure Claude Desktop as follows in your claude_desktop_config.yml:

{
  "mcpServers": {
    "embabel": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "http://localhost:8080/sse"
      ]
    }
  }
}

See MCP Quickstart for Claude Desktop Users for how to configure Claude Desktop.

The MCP Inspector is a helpful tool for interacting with your Embabel SSE server, manually invoking tools and checking the exposed prompts and resources.

Start the MCP Inspector with:

npx @modelcontextprotocol/inspector

Consuming MCP Servers

The Embabel Agent Framework provides built-in support for consuming Model Context Protocol (MCP) servers, allowing you to extend your applications with powerful AI capabilities through standardized interfaces.

What is MCP?

Model Context Protocol (MCP) is an open protocol that standardizes how applications provide context and extra functionality to large language models. Introduced by Anthropic, MCP has emerged as the de facto standard for connecting AI agents to tools, functioning as a client-server protocol where:

Clients (like Embabel Agent) send requests to servers
Servers process those requests to deliver necessary context to the AI model

MCP simplifies integration between AI applications and external tools, transforming an "M×N problem" into an "M+N problem" through standardization - similar to what USB did for hardware peripherals.

Configuring MCP in Embabel Agent

To configure MCP servers in your Embabel Agent application, add the following to your application.yml:

spring:
  ai:
    mcp:
      client:
        enabled: true
        name: embabel
        version: 1.0.0
        request-timeout: 30s
        type: SYNC
        stdio:
          connections:
            docker-mcp:
              command: docker
              args:
                - run
                - -i
                - --rm
                - alpine/socat
                - STDIO
                - TCP:host.docker.internal:8811

This configuration sets up an MCP client that connects to a Docker-based MCP server. The connection uses STDIO transport through Docker's socat utility to connect to a TCP endpoint.

Docker Desktop MCP Integration

Docker has embraced MCP with their Docker MCP Catalog and Toolkit, which provides:

Centralized Discovery - A trusted hub for discovering MCP tools integrated into Docker Hub
Containerized Deployment - Run MCP servers as containers without complex setup
Secure Credential Management - Centralized, encrypted credential handling
Built-in Security - Sandbox isolation and permissions management

The Docker MCP ecosystem includes over 100 verified tools from partners like Stripe, Elastic, Neo4j, and more, all accessible through Docker's infrastructure.

Learn More

A2A

Embabel integrates with the A2A protocol, allowing you to connect to other A2A-enabled agents and services.

Enable the a2a Spring profile to start the A2A server.

You'll need the following environment variable:

GOOGLE_STUDIO_API_KEY: Your Google Studio API key, which is used for Gemini.

Start the Google A2A web interface using the a2a Docker profile:

docker compose --profile a2a up

Go to the web interface running within the container at http://localhost:12000/.

Connect to your agent at host.docker.internal:8080/a2a. Note that localhost:8080/a2a won't work as the server cannot access it when running in a Docker container.

Running Tests

Run the tests via Maven.

mvn test

This will run both unit and integration tests but will not require an internet connection or any external services.

Spring profiles

Spring profiles are used to configure the application for different environments and behaviors.

Interaction profiles:

shell: Runs agent in interactive shell. Does not start web process.

Model profiles:

ollama: Looks for Ollama models. You will need to have Ollama installed locally and the relevant models pulled.
docker-desktop: Looks for Docker-managed local models when running outside Docker but talking to Docker Desktop with the MCP extension. This is recommended for the best experience, with Docker-provided web tools.
docker: Looks for Docker-managed local models when running in a Docker container.

Logging profiles:

severance: Severance specific logging. Praise Kier!
starwars: Star Wars specific logging. Feel the force
colossus: Colossus specific logging. The Forbin Project.

Testing

A key goal of this framework is ease of testing. Just as Spring eased testing of early enterprise Java applications, this framework facilitates testing of AI applications.

Types of testing:

Unit tests: All agents are unit testable, like any Spring-managed beans. Construct them with mock objects; call individual action methods. The testing library facilitates testing prompts.
Integration tests: tbd

Logging

All logging in this project is either debug logging in the relevant class itself, or results from the stream of events of type AgentEvent.

Edit application.yml if you want to see debug logging from the relevant classes and packages.

Available logging experiences:

severance: Severance logging. Praise Kier
starwars: Star Wars logging. Feel the force. The default as it's understood throughout the galaxy.
colossus: Colossus logging. The Forbin Project.
montypython: Monty Python logging. No one expects it.
hh: Hitchhiker's Guide to the Galaxy logging. The answer is 42.

If none of these profiles is chosen, Embabel will use vanilla logging. This makes me sad.

Adding Embabel Agent Framework to Your Project

Maven

The easiest way is to add the Embabel Spring Boot starter dependency to your pom.xml:


<dependency>
    <groupId>com.embabel.agent</groupId>
    <artifactId>embabel-agent-starter</artifactId>
    <version>${embabel-agent.version}</version>
</dependency>

Gradle (Kotlin DSL)

Add the required repositories to your build.gradle.kts:

repositories {
    mavenCentral()
    maven {
        name = "embabel-snapshots"
        url = uri("https://repo.embabel.com/artifactory/libs-snapshot")
        mavenContent {
            snapshotsOnly()
        }
    }
    maven {
        name = "Spring Milestones"
        url = uri("https://repo.spring.io/milestone")
    }
}

Add the Embabel Agent starter:

dependencies {
    implementation('com.embabel.agent:embabel-agent-starter:${embabel-agent.version}')
}

Gradle (Groovy DSL)

Add the required repositories to your build.gradle:

repositories {
    mavenCentral()
    maven {
        name = 'embabel-snapshots'
        url = 'https://repo.embabel.com/artifactory/libs-snapshot'
        mavenContent {
            snapshotsOnly()
        }
    }
    maven {
        name = 'Spring Milestones'
        url = 'https://repo.spring.io/milestone'
    }
}