If Generative AI is the Brain, Agentic AI is the Acting Agent | TechBlog | Research

Key Points Summary

Samsung SDS AI Research Team - The Samsung SDS AI Research Team is composed of four labs focusing on reinforcement learning, generative AI, agents, and advanced technologies. Through collaborative research, the team develops AI solutions tailored for enterprise environments.
AI Solutions Optimized for Enterprise Environments – With over 40 years of expertise in enterprise digital transformation, Samsung SDS delivers solutions that effectively integrate AI technologies with existing legacy systems. Their practical approach prioritizes achieving the best results with optimal technology.
Evolution from Generative AI to Agentic AI - The future of AI is defined by Agentic AI. Agentic AI is an autonomous artificial intelligence capable of independently planning and executing problem-solving tasks. As explained by Tae-hee Lee, generative AI serves as the 'brain,' creating information, while Agentic AI acts as the 'agent,' executing actions.

We had the opportunity to meet Tae-hee Lee, Head of AI Research Team at Samsung SDS Research. After earning his Ph.D. in computer vision from UCLA, Tae-hee Lee led a photo search project at Google. Currently, he is spearheading the development of Agentic AI and hyper-automation technologies at Samsung SDS Research. Recently, Tae-hee Lee has presented "The Paradigm of Agentic AI beyond Generative AI" at several external events, including REAL 2024. Now, let us introduce the future that the AI Research Team at Samsung SDS is poised to build, as shared by Tae-hee Lee.

"For me, a computer was a playground for solving problems"

What experiences and journey have brought you to where you are today?

Continuous challenges have been the greatest driving force in my life. After graduating with a bachelor's degree in School of Computing from KAIST, I decided to pursue further studies in the United States. I completed my master's at UCSB and my Ph.D. at UCLA. Since childhood, computers have always felt like a playground to me - solving problems was fun. Building on this interest, I continued to embrace challenges, which ultimately led me to the field of "computer vision”.

The period from 2007 to 2012, when I was conducting research at UCLA, marked a pivotal moment in the field of object recognition technology, as it transitioned from traditional computer vision methods to deep learning. I had the opportunity to experience and participate in this paradigm shift in computer technology firsthand. My research led to a demo of moving object recognition on mobile phones, which caught the attention of a Google representative. This resulted in an internship offer, which eventually led to my joining Google as an engineer. At Google, I primarily worked on developing photo search functionality. I was able to witness the process of this feature being implemented into products and utilized by users worldwide.

"Best technology vs. Optimal technology"

What is the biggest lesson from your experience at Google that still influences you today?

The biggest lesson is learning to balance technical excellence with practical applicability. There’s a difference between the research lab and reality. An AI model that performs exceptionally well in the research lab may face numerous real-world constraints when applied to actual services. At Google, I learned to recognize and bridge this gap. Through the process of integrating image recognition technology into real-world products, I realized that technically superior technology is not always the best for users or businesses. I developed the insight to choose not the theoretically best technology, but the optimal one that can solve current problems.

This experience also helps me in developing enterprise AI solutions at Samsung SDS. While I primarily worked on B2C products at Google, the products I develop at Samsung SDS are B2B services used in corporate environments. These require careful consideration of factors like stability, security, and compatibility with existing systems—making it even more important to recognize these differences. In this intersection of technology and real-life applications, I strive to develop AI that creates tangible business value, rather than technology for its own sake.

It seems you're developing not just the best technology, but the optimal one! To provide the optimal technology, there must be many considerations and decisions involved. Do you have any principles for decision-making as Head of AI Research Team?

In an environment where AI technology is rapidly advancing, it's impossible to research every field. Therefore, it's crucial to identify areas where we can achieve the maximum impact with our existing capabilities and resources. This involves setting priorities. Especially with the emergence of innovative technologies like GPT and generative AI, there have been many moments when we had to reevaluate our research priorities. Specifically, in late 2022 after joining Samsung SDS, one of our researcher suggested, "We should focus on generative AI next year," before the ChatGPT boom. Since generative AI requires significant investment, setting research priorities involved many considerations, but our Samsung SDS Research seemed to sense this change early on. In fact, from November when this suggestion was made, the following year flew by so quickly that we barely had time to flip the calendar. (Laughs) Samsung SDS Research, Development Teams and Business Teams all worked together seamlessly as one and quickly adapted to the changes after the emergence of generative AI technology. As a result, Samsung SDS’s generative AI platform, FabriX, was born.

What is Samsung SDS FabriX?

Samsung SDS’s generative AI platform, FabriX, is an enterprise generative AI solution that connects enterprise knowledge - both internal and external - and various systems like ERP, SCM, MES, and CRM with generative AI to drive innovation in business automation and productivity.

When setting research priorities, it is essential to listen to the ideas of researchers. Reflecting on past experiences, the directions proposed by researchers have often proven to be remarkably accurate.

For instance, earlier this year, we reviewed the "Byte Latent Transformer" research published by Meta to guide our decision-making process. To evaluate the potential impact of this technology in advance, we conducted intensive experiments with researchers over approximately a month and a half. Our analysis concluded that the technology was still in its nascent stage. Based on this assessment, we established future research priorities and directions. This approach enabled us to thoroughly understand the potential and limitations of the technology with minimal investment. Notably, this judgment has remained valid even several months later.

Whenever possible, I strive to fully listen to the ideas of each individual team member and verify their ideas through small-scale experiments. This principle has been instrumental in my decision-making process.

" The Samsung SDS AI Research Team conducts research that complements individual expertise "

What are the labs that make up the Samsung SDS AI Research Team you are part of?

The Samsung SDS AI Research Team is composed of four specialized labs. These include the Reinforcement Learning Lab, which focuses on reinforcement learning; the Gen.AI Core Lab, which researches generative AI models; the AI Advanced Research Lab, which explores next-generation technologies such as RAG; and the Autonomous Intelligence Lab, which investigates Agentic AI. These labs collaborate in a mutually complementary manner, leveraging their respective expertise to drive innovation.

Yeong-dae Kwon, the leader of the Reinforcement Learning Lab, has achieved remarkable success in his research, including presenting papers at NeurIPS (Neural Information Processing Systems) for three consecutive years. His work on "Pomo: Policy optimization with multiple optima for reinforcement learning (2022)," which address multi-optimization challenges in reinforcement learning, garnered significant academic attention. Currently, Lab Leader GWON Yeongdae is working on a research project involving a GUI-using Agent.

What is Reinforcement Learning?

Reinforcement learning is a machine learning methodology where an Agent learns an optimal action policy by interacting with an environment to maximize rewards. The Agent selects actions, and the environment responds by providing a new state and a reward, creating a cyclical feedback loop. Recently, Multi-Agent Reinforcement Learning (MARL) and meta-reinforcement learning have emerged as prominent areas of focus. Samsung SDS has made notable contributions to this field, including the publication of a research paper titled "Reinforcement Learning for Combinatorial Optimization" at NeurIPS, one of the world's most prestigious artificial intelligence conferences, and receiving the Jury Prize from the conference panel.

Furthermore, the Autonomous Intelligence Lab is currently focusing on developing "Software Engineering Agents." For Samsung SDS, software development and operation are central to its business and research efforts. The lab is actively working on automating Agents designed to boost developer productivity. These include developing models that excel in coding, automatically generating tests, and even creating technologies to automatically generate development documents from existing code, thereby providing comprehensive support for software development processes.

"Agentic AI is an acting artificial intelligence"

What is the core technology that Samsung SDS AI Research Team is most focused on?

The core technology we are most focused on is Agentic AI. The Samsung SDS AI Research Team is dedicated to researching how Agentic AI can drive hyper-automation for enterprises, extending beyond the capabilities of simple generative AI, and applying these advancements to real-world projects.

What is Agentic AI?

Agentic AIis an AI system that acts and makes decisions in a goal-oriented manner, going beyond merely generating information. It autonomously performs a series of processes, including analyzing problems, planning, executing actions, and evaluating results, by leveraging reasoning techniques such as ReAct (Reasoning and Acting), Tree of Thoughts (ToT), and Chain of Thought (CoT). While generative AI serves as a "creative tool" that generates content based on large-scale learned knowledge in response to user instructions, Agentic AI acts as an "autonomous problem solver" that independently plans, executes, and resolves issues to achieve its objectives.

What is the primary difference between Generative AI and Agentic AI?

Generative AI is focused on creating content, such as text and images. In contrast, Agentic AI goes beyond creation, emphasizing the ability to perform actions autonomously to solve problems.

To simplify, if generative AI is akin to a brain that generates information, Agentic AI can be likened to an entity that has a brain but also arms and legs to takes actions. In other words, Agentic AI not only plans what needs to be done, but also executes the plan, evaluates the results, delivers outcomes to the user, overseeing the entire process of action.

Compared to traditional automation technology like RPA (Robotic Process Automation), the differences are more pronounced. While RPA focuses on automating predefined repetitive tasks, Agentic AI can autonomously plan and execute tasks even when they are not predefined, given a specific goal. It can be regarded as an advanced form of RPA enhanced with the capabilities of generative AI.

Recent advancements in technologies like MCP (Model Context Protocol) have further expanded the potential of Agentic AI. For Agentic AI to function effectively, it requires channels to communicate with various systems, databases, and even other Agents. Protocols like MCP facilitate seamless communication, enabling Agentic AI to operate efficiently.

What changes do you expect Agentic AI to bring to the corporate environment?

When fully introduced, Agentic AI is expected to transform the corporate environment by evolving beyond simple task automation. It will establish a collaborative model where employees and AI work together in complex and comprehensive business processes. Tasks that employees transitionally handled from start to finish will increasingly be autonomously managed by Agents, allowing employees to shift their focus toward providing higher-level feedback or strategic guidance.

At Samsung SDS, several Agentic AI cases are already under development with notable example being the software development Agent mentioned earlier. This Agent not only writes code or generates tests but also supports document creation. This Agent is currently utilized within Samsung SDS to minimize potential human errors in software development processes, thereby enhancing both development speed and quality.

Another example is the time-series data analysis Agent. This Agent analyzes log data generated during IT system operations to predict or diagnose system failures. The technology behind this Agent can be applied across various industries, such as risk analysis for financial companies or predicting delivery arrival times in logistics, making it highly versatile.

What should a company prepare if it wants to adopt Agentic AI?

For effective adoption of Agentic AI, it is essential to establish a well-structured data environment. Just as a work environment and space are important for humans, a robust data environment is a critical factor for Agents. Therefore, companies need a platform that manages data in a way that enables Agents to utilize it efficiently and securely. Through this, a system must be established that allows Agents to access and leverage corporate data effectively.

"We provide solutions for AI adoption in enterprise environments"

It seems there are differences between how individuals and companies use AI. What are the main differences?

Many companies face substantial challenges when adopting AI, particularly in effectively integrating new AI technologies with existing legacy systems. These systems, which have evolved over years or even decades, cannot simply be replaced with the latest technology.

Samsung SDS leverages its 40+ years of expertise in corporate digital transformation to address this issue. The key is ensuring that AI adoption doesn’t end as a one-time project but continues to deliver ongoing value by establishing a robust connection to actual business operations. This is where Samsung SDS excels.

For hyper-automation in enterprises, Samsung SDS offers three key solutions. First, FabriX is a generative AI platform that enables users to easily create and share Agents using the Agent Studio development tool. Brity Copilot is a personalized AI Agent that provides tailored support by leveraging the user's data. Lastly, Brity Automation offers workflow solution that automates entire IT and business processes, rather than individual tasks.

In particular, FabriX serves as a platform that bridges the gap between legacy systems and new generative AI technologies. The AI research team collaborates closely with the corresponding business team from the planning stage to develop products.

The ultimate goal of Samsung SDS’s AI Research Team is to move beyond simple chatbots and AI Copilots and achieve Autonomous Agents - Agentic AI capable of autonomously making decisions and taking action to solve corporate problems. This involves advancing toward a stage where AI can make judgments and act independently with minimal user intervention.

"For continuous growth, we must focus on defining problems"

What capabilities are essential for companies and researchers to sustain growth in the rapidly evolving AI technology environment today?

I believe that the most critical focus for both companies and researchers right now is to ask themselves, What problem should we solve? For companies it’s essential to identify urgent issues that need addressing and assess their long-term relevance.

Researchers face a similar challenge. They must evaluate the present and future impact of the problems they tackle. If a problem is significant, it’s likely others are working on it too. In such cases, continuous engagement—whether through collaboration or competition—is vital. While this process can be exhausting or even disheartening at times, witnessing dynamic change as an engineer in this era is both exhilarating and deeply rewarding.

In particular, the ability to quickly recognize technology trends is crucial. In an environment where countless papers and technical blogs are published daily, developing the discernment to identify truly impactful technologies is essential. While it’s easy to get lost in the vast amount of information, the effort to extract the core insights is necessary. Recently, Agentic AI has proven invaluable in organizing and analyzing these technologies effectively.

When introducing AI technology in a corporate setting, ensuring it becomes a sustainable competitive advantage rather than a fleeting trend is paramount. Rather than merely adopting the latest technology, it’s essential to continuously validate and refine how that technology can create lasting value for the business.

The AI field is currently experiencing one of its most exciting periods. Numerous problems are being solved, and new technologies are emerging at an unprecedented pace. To thrive in this environment, a mindset of continuous learning, self-challenge, and growth through failure is essential.

"Where research meets reality: The Samsung SDS AI Research Team"

What would you like to say to AI researchers who are interested in joining the Samsung SDS AI Research Team?

The greatest strength of the Samsung SDS AI Research Team lies in its environment where research outcomes can be immediately integrated into business and real-world applications.

At the Samsung SDS AI Research Team, you can experience the full lifecycle of software development and its application to business, beyond merely publishing papers. The ability to directly connect theoretical research with practical implementation is what makes Samsung SDS Research particularly compelling.

I’d also like to highlight that our work focuses on solving problems for enterprise, not individuals. In fact, developing technologies that address corporate-level problems can be quite challenging. However, this "challenge" can be a significant draw for engineers like me who thrive on tackling complex problems. (laughs)

Additionally, the process of discovering and solving important problems is highly dynamic, as researchers with diverse backgrounds come together. At Samsung SDS Research, the AI Research Team does not work alone; it collaborates synergistically with other teams such as the Cloud Research Team and the Security Research Team, allowing researchers to gain exposure to a broad spectrum of IT technologies beyond AI.

Lastly, unlike other general generative AI, as Agentic AI emphasizes "defining the problem" as its starting point, rather than merely generating solutions, Samsung SDS Research seeks individuals who can excel in this critical phase of problem definition, rather than solely focusing on solving predefined challenges. Moving forward, we are dedicated to advancing AI technologies that can evolve sustainably in the dynamic technological and corporate landscapes, including innovations like Agentic AI and beyond.

Now, Samsung SDS is transitioning from Korea’s largest SI company into a leading cloud provider, offering a diverse range of services. For those eager to be part of this transformation, the AI Research Team at Samsung SDS Research welcomes you with open doors.

"If Generative AI is the Brain, Agentic AI is the Acting Agent"