Java Developers Get Multiple Paths To Building Ai Agents

Earlier this year, I wrote an article astir really Java would soon overtake Python arsenic nan programming connection of choice for AI app development.

Back then, Simon Ritter, lawman CTO at Azul Systems, told maine that based connected Azul research, including a survey of Java developers, Java could encroach on Python’s lead successful usage for AI improvement wrong a twelvemonth and a half. In fact, Ritter said 2025 would beryllium nan past twelvemonth of Python’s power for AI development.

So what tin we expect successful 2026? In nan mediate of my investigation connected nan topic, I noticed that my aged pal James Governor of RedMonk had written a potent portion astir Java and agentic systems development. Then conscionable nan different day, connected our ain TNS site, Michael Coté, different astute expert I know, wrote a portion entitled “Your Enterprise AI Strategy Must Start With Java, Not Python.” So, I knew I was connected nan correct track.

“It’s good known successful developer circles that Java is amended for processing endeavor AI applications, fixed amended scalability and performance, but correct now Python outpaces Java pinch its libraries and different infrastructure to support nan improvement of AI,” Ritter told The New Stack successful a caller interview. “However, enterprises are realizing that Java is nan amended prime for enterprise-level deployments. We’re apt to spot Java outpace Python wrong nan adjacent 18 months to 3 years.”

Java successful nan Agentic AI Era

Now, arsenic we are successful nan agentic AI era, respective Java AI supplier frameworks and libraries — specified arsenic Embabel, Koog, LangChain4j and others — person emerged to support Java developers building AI applications.

Yet nan mobility remains whether Java tin drawback up to Python successful AI development.

“My individual sentiment is that nan expanding readiness of Java frameworks will thief adjacent nan spread betwixt Python and Java usage and successful nan AI supplier space,” Azul’s Ritter told The New Stack successful a caller interview. “Although Python is simply a very celebrated connection for AI development, halfway technologies, specified arsenic LLMs [large connection models] and Torch, are written chiefly successful C and C++. Python has go celebrated owed to nan wide readiness of libraries for illustration PyTorch and its perceived easiness of use.”

However, “As much Java frameworks and libraries are developed — specified arsenic LangChain4j — I deliberation developers will find that Java offers additional, much compelling advantages,” Ritter said. “Java excels astatine scalability and performance, particularly successful handling aggregate concurrent threads of execution (something Python struggles with). With nan exponential summation successful usage of AI agents, factors specified arsenic capacity will go progressively important.”

A Fighting Chance

Java has a “chance” successful this undeclared battle. Some opportunity much than conscionable a chance.

“Java has a chance to drawback up to Python for AI apps astatine nan endeavor eventually, but I don’t deliberation that will hap tomorrow,” Andrew Cornwall, an expert astatine Forrester Research, told The New Stack.

But “AI researchers are still much comfortable successful Python than Java, truthful Python will spot nan bleeding separator investigation first,” he noted. “However, enterprises find Java easier to integrate. Java’s sewage a much robust unreality ecosystem and much mature tooling. Right now, nan manufacture is struggling to support up pinch GenAI [generative AI] advances, truthful building pinch a dynamically typed connection for illustration Python is faster. Once nan gait settles down and patterns go good known, frameworks for Java will beryllium available, and astir apt much desirable for enterprises.”

Brad Shimmin, an expert astatine The Futurum Group, said he does not deliberation that Java will overtake, aliases “displace” Python arsenic nan preferred connection for doing information science.

“Honestly, Haskell and Mojo — an hold of Python — guidelines a amended chance of that. As we’ve seen clip and again, developers prime a) what they cognize and b) what they deliberation will get nan occupation done astir effectively,” he said. “I deliberation much developers see building AI outcomes successful Python than successful Java, simply owed to relation and familiarity. You tin spot this if you comparison GitHub projects for supportive libraries. Right now, LangChain has astir 1,132% much stars than LangChain4j. Does that make Python better? Of people not. I deliberation having devices for illustration LangChain4j and Crew4J astatine nan fresh is critical, fixed nan number of developers penning mission-critical codification successful Java. Frankly, a batch of Python codification gets refactored into Java — and Scala — successful bid to present mission-critical performance, security, etc.”

But we’re talking astir nan enterprise.

Asked if he believed Java could flooded Python for AI app development, Dmytro Liubarskyi, LangChain4j creator and Red Hat elder package engineer, replied: “When it comes to integrating LLMs into enterprise applications, I judge it can. Python will apt stay nan ascendant connection for research, experimentation, and prototyping. However, Java is exceptionally well-positioned for accumulation usage successful endeavor environments.”

Moreover, Liubarskyi added: “Java has agelong been nan de facto modular for building large-scale endeavor systems, and I’m assured it will besides go nan connection of prime for endeavor AI and LLM integration.”

Embabel: Targeting nan JVM and nan Enterprise

Rod Johnson, nan creator of Embabel and besides nan creator of nan wide utilized Spring Framework, said he believes Java is amended suited to building AI apps successful nan enterprise.

Johnson told The New Stack that he had been doing a batch of instrumentality learning (ML) stuff, chiefly successful Python, utilizing TensorFlow and PyTorch, and training neural networks.

“So, I decided that it’d beryllium absorbing to spot what nan perfect model would look for illustration for building agents. And I reasonably quickly concluded that for what I wanted to do, and what I deliberation is unlocking nan worth of existing businesses, nan JVM [Java Virtual Machine] is going to beryllium a overmuch amended spot to do that, because you want to beryllium capable to turn that functionality retired of nan valuable halfway business functionality you’ve already got, alternatively than adhd thing that’s alien, that conscionable talks to it,” he said.

Johnson added that “Python’s awesome for information science, but you’re not talking astir that erstwhile you’re GenAI enabling business applications. You’re talking astir exertion development, accepted exertion development, alternatively than information subject aliases axenic ML, and Java has ever been astir apt nan azygous astir celebrated connection for building those kinds of applications, and there’s a batch of logic down that. So therefore, to me, it makes a batch of consciousness to build it connected nan JVM and build it connected apical of Spring, build it connected apical of nan functionality and integrations that group already have, and create thing that’s really easy for developers to adopt.”

Spring Boot Advantage

Indeed, Johnson said if a developer understands Spring Boot, they’re going to beryllium up and moving pinch Embabel successful nether 5 minutes.

Embabel is an supplier framework, initially for nan JVM, that is intended not conscionable to play catch-up pinch Python supplier frameworks, but to leapfrog them, Johnson has said successful blog posts.

Johnson told The New Stack that arsenic this abstraction matures, group recognize that nan cardinal adjacency is not calling your LLM aliases information science. The cardinal adjacency is your existing business functionality and your existing developer accomplishment set. The existing developer accomplishment group successful endeavor is Java, he noted.

“So, erstwhile you deliberation astir it that way, successful a GenAI application, you’re calling nan LLM complete HTTP. It’s not successful process. You’re not utilizing thing magical that Python has to do that. It’s typically an HTTP telephone that’s truthful simple,” Johnson said.

Better Programming Model

Johnson primitively created the Spring Framework to present a amended programming exemplary for endeavor Java, and now Embabel is aimed astatine doing nan aforesaid for building JVM-based AI agents.

Johnson explained that, successful Python, you create a authorities machine; successful Embabel, you deploy your actions, and nan model tin activity retired what bid to tally those actions successful to execute nan goal. It’s nan model that’s doing it, utilizing a readying attack called goal-oriented action planning.

“And nan beauty of it is that it’s some smart, but it’s deterministic, truthful nan model will beryllium capable to show you why it did that, and if it finds itself pinch nan aforesaid input objects, again, it will do nan aforesaid thing,” he said. That is “definitely a valuable differentiator successful enterprise, because explainability is critical,” Johnson said. “Determinism is critical, and I judge Embabel is nan only model that is perfectly straight focused connected that while retaining nan expertise to beryllium smart, and for illustration you adhd much actions and tin do much things that you didn’t explicitly program.”

Why Kotlin

Embabel is written successful Kotlin, a modern, statically typed programming connection from JetBrains, known for its conciseness and interoperability pinch Java arsenic it targets nan JVM.

“I personally for illustration programming successful Kotlin to programming successful Java,” Johnson said. Moreover, “we consciousness our velocity successful processing nan model will beryllium faster successful Kotlin alternatively than Java, but we person written astir of our sample applications and rather a batch of our soul applications successful Java.”

JetBrains Koog

Koog is JetBrains’ Kotlin-native model for creating AI agents that tally locally, interact pinch devices and automate analyzable tasks. JetBrains created Kotlin arsenic a JVM-based replacement to Java that is known for its conciseness, null information features and afloat interoperability pinch Java.

“For galore enterprises, Python is not considered a production-ready language, moreover though astir modern AI devices are built connected it,” Vadim Briliantov, tech lead for Koog astatine JetBrains, told The New Stack. “That puts developers who usually activity successful type-safe languages for illustration Java aliases Kotlin successful a reliable spot. Koog efficaciously bridges this gap, offering these developers a long-needed solution, and abstracts a batch of nan ‘AI glue code’ distant from devs who don’t needfully travel each nan ML lit successful a type-safe way.”

Meanwhile, Briliantov echoes what Johnson said astir Python versus Java regarding LLM-based AI.

“While it mightiness look counterintuitive, Python’s advantage successful ML libraries and devices doesn’t really use to LLM-based AI,” he said. “An LLM telephone is fundamentally conscionable an HTTP petition to immoderate outer service. No magic here. At nan aforesaid time, an AI supplier is simply a long-running strategy that mightiness neglect aliases beryllium restarted connected different machine. A strategy that interacts pinch your database and authorizes immoderate personification actions.”

Real-World Advantages

Briliantov asserts that Java has important advantages complete Python for “real-world” applications.

“People person been building specified systems for decades connected nan JVM and for a bully reason. Once AI demos time off nan lab, they look real-world limitations,” he said. “They request to beryllium fault-tolerant, for example. And Koog already provides specified responsibility tolerance pinch database integrations.”

JVM Advantages

The JVM delivers different important advantages complete Python, Briliantov said.

The first advantage is its integration into nan existing endeavor ecosystem. If your endeavor backend is already connected nan JVM, there’s nary logic to move nan stack for nan liking of AI, he argues.

The 2nd advantage is that its type information and domain modeling not only let you to build much maintainable code, but besides much predictable and amended AI agents, he noted.

“Instead of relying connected punctual engineering everywhere, you tin picture what you want arsenic a information class; Koog would guarantee that an agentic measurement would nutrient a consequence that complies pinch nan shape — thing tin beryllium occasionally forgotten by nan LLM. It’s for illustration a paradigm shift: You only deliberation astir nan information and nan process, and nan model enforces that nan LLM follows your type-safe rules,” Briliantov said.

Crew4J

Crew4J is simply a Java-based model for building and managing collaborative multiagent AI systems. It is not related to CrewAI.

“My ngo is to make precocious AI capabilities accessible to Java developers and enterprises without forcing them to wantonness their existing infrastructure and expertise,” Mahesh Awasare, nan creator of Crew4J, said successful a connection astir his improvement philosophy.

Awasare is an endeavor designer moving astatine Avaloq, a financial merchandise company. He has besides worked for different world banks and financial institutions specified arsenic HSBC, Citi, Credit Suisse and Worldline.

“I noticed that nan endeavor systems are written successful Java and 90% of endeavor systems are very difficult to migrate to a Python tech stack to leverage AI,” Awasare told The New Stack. “So, I thought, why not create thing that will beryllium easy for existing implementations to adopt for AI agentic development. This is nan logic I created Crew4J.”

He said he believes that “all endeavor systems are looking to merge AI, and since astir of them are successful Java, nan request for specified a model is high. People are trying to operation Java implementations pinch Python for integrating AI, and it is not easy.”

However, Crew4J is very elemental to use, requiring conscionable a six-step process.

Still, Awasare said he does not judge Java will outpace Python for AI development. “I americium not judge if it would overtake, but would it co-exist? BIG YES,” he responded via email.

LangChain4j

LangChain4j is an unfastened root Java model that simplifies nan integration of LLMs and different AI capabilities into Java applications.

It provides unified APIs for a wide scope of LLM providers and vector stores, making it easy to move betwixt them pinch minimal codification changes.

The room besides includes galore out-of-the-box building blocks for Retrieval-Augmented Generation (RAG) pipelines, agents and related patterns, allowing developers to attraction connected exertion logic alternatively of boilerplate, LangChain4j creator Liubarskyi told The New Stack.

It is designed specifically for Java developers who want to build production-ready LLM applications and agentic systems.

How It All Started

In early 2023, Liubarskyi was moving connected LLM-powered chatbots successful Java and quickly realized that location were nary suitable libraries available. Everything had to beryllium implemented from scratch. Around that time, he discovered LangChain and thought, “We request thing for illustration this, but for Java.” That thought became nan instauration of LangChain4j and moreover inspired nan name.

Despite nan name, LangChain4j is not a nonstop larboard of LangChain, he said. “Instead, it combines ideas and concepts from LangChain, Haystack, LlamaIndex, and nan broader community, on pinch our ain innovations, each adapted to Java’s ecosystem and creation principles,” Liubarskyi told The New Stack.

The request has been very strong. After LangChain4j was presented astatine nan Devoxx Belgium conference successful nan autumn of 2023, nan task gained important momentum and has continued to turn steadily crossed each metrics. It has precocious reached 10,000 GitHub stars and shows accordant maturation successful some downloads and unsocial users.

The organization is ample and active, pinch a changeless watercourse of propulsion requests and contributions. Today, LangChain4j is utilized by a wide scope of teams, from startups to ample enterprises.

Indeed, nan sermon astir LangChain4j attracted nan attraction of Microsoft, which has invested successful nan task and helped to heighten information for it.

Microsoft Investment

In a blog post entitled “Microsoft and LangChain4j: A Partnership for Secure, Enterprise-Grade Java AI Applications,” Julien Dubois, main manager, Java developer relations, wrote: “Our telemetry information indicates important maturation successful LangChain4j adoption, pinch hundreds of Microsoft customers utilizing nan model successful accumulation environments. Recognizing nan value of this open-source initiative, Microsoft has made important contributions to heighten nan project’s Azure and OpenAI integration capabilities.”

In a statement, Liubarskyi said: “We’re particularly grateful for Microsoft’s dedication to strengthening nan information of nan room — their adjacent collaboration and thorough audit process person helped america build a much unafraid and trustworthy instauration for AI improvement successful Java.”

Distinct From nan Alternatives

Meanwhile, Liubarskyi said LangChain4j is chopped from Embabel, Koog and nan others successful respective different ways.

He said his extremity was to support LangChain4j lightweight yet powerful. It is intentionally designed arsenic an unopinionated room alternatively than a afloat framework, which makes it easy to usage pinch immoderate JVM connection and merge into virtually immoderate stack.

“It useful seamlessly pinch celebrated Java frameworks and runtimes specified arsenic Quarkus, Spring, Helidon, Micronaut, WildFly and Liberty,” he said.

LangChain4j was besides 1 of nan first libraries successful this space, which intends it has had clip to mature.

“We spot a beardown accent connected API stableness and effort to debar breaking changes arsenic overmuch arsenic possible,” he added. “Another cardinal spot is nan breadth of integrations: Out of nan box, users tin activity pinch a wide assortment of LLM providers and vector stores.

“Finally, nan room offers 2 chopped levels of abstraction. Developers tin usage low-level APIs for maximum control, aliases higher-level AI services and agents to get much functionality pinch little setup.”

Group Created pinch Sketch.