[AWS re:Invent] "Self-thinking, Predicting, and Executing 'AI Agents' Will Be a Turning Point for Industry"

Day 3 of re:Invent: Keynote Speech by Vice President Subramanian
New Features Introduced for AI Model Development Platform 'SageMaker'
Teacher Model Now Trains Student Model with Innovative Capabilities
AI Model Enhanced Through Self-Learning and Reinforcement in a 'Gym'

"AI agents are transforming a wide range of industries, from software development and drug discovery to precision agriculture and architectural design."

Swami Subramanian, Vice President of Authentic AI at Amazon Web Services (AWS), stated during his keynote speech on "The Future of Authentic AI" held on the morning of the 3rd (local time) at the Venetian Hotel in Las Vegas, Nevada, USA, that "the emergence of AI agents is accelerating the pace of change in industry."

On the third day of AWS's largest annual event, "re:Invent 2025," AI agents took center stage. This focus comes as self-planning and self-executing AI agents are expected to become a turning point for many industries. On this day, AWS unveiled new features for the development of AI agents and for strengthening AI models.

Swami Subramanian, Vice President of AWS Authentic AI, delivered a keynote speech on "The Future of Authentic AI" on the morning of the 3rd (local time) at the Venetian Hotel in Las Vegas, Nevada, USA. Provided by AWS

AI agents go beyond existing AI assistants, which only provided limited support for users' work. Now, AI can think and predict on its own, and even perform tasks on behalf of the user. For example, while traditional AI assistants would only search for flights that meet certain conditions when a user requests a flight booking, AI agents can actually make the reservation and complete the payment. Subramanian explained, "When an AI assistant receives a user's question, it tells you what to look up, but an AI agent directly investigates, diagnoses, and offers solutions."

AI agents are also enhancing the efficiency of developers' work. Subramanian added, "With the help of AI agents, we no longer need to memorize programming language syntax or application programming interface (API) calls," and "tasks that used to take years or months in the past can now be completed in just weeks or even days."

During his keynote, Subramanian introduced new features for "SageMaker," a cloud platform for AI model development. Specifically, features such as Supervised Fine Tuning, Reinforcement Fine Tuning, and Customization for AI models were showcased. SageMaker helps developers build and train AI models on the cloud.

Among these, supervised fine tuning involves fine-tuning a model for its intended purpose through mutual learning between a teacher model and a student model. High-performing, large-scale teacher models transfer knowledge to smaller, lighter student models through a process called distillation. Distillation refers to the technology of transferring the knowledge from large, complex AI models to smaller, more efficient ones.

The "Amazon Nova Act," an AI agent model that operates in web browsers, was also introduced. This model analyzes websites and performs tasks on behalf of users, such as document creation, replying to emails, shopping, and restaurant reservations. Subramanian explained, "We improved Nova Act through reinforcement learning," adding, "We created a kind of 'gym' where the agent learns and trains-receiving rewards for successful interactions and penalties for failures."

Swami Sivasubramanian, Vice President of AWS Authentic AI, delivered a keynote speech on "The Future of Authentic AI" on the morning of the 3rd (local time) at the Venetian Hotel in Las Vegas, Nevada, USA. Provided by AWS

Subramanian defined the key components of AI agents as models, code, and tools, emphasizing, "The real power lies in how all these components are orchestrated." He noted, "In the past, developers had to connect these three components one by one through coding, which required a lot of time." He continued, "Now, as AI agents autonomously plan actions and connect tools, they can adapt to unpredictable situations. Developers only need to define the components such as models, code, and tools."

Additionally, an episodic memory function has been added to the Agent Core. This feature allows AI agents to learn from past experiences and apply this knowledge to future interactions. AI agent capabilities have also been added to "Amazon Connect," a contact center solution that can be used in customer service centers.

Subramanian concluded his presentation by saying, "With AI agents, we will be able to solve problems that once seemed impossible and enjoy the freedom to create without limits every day."

Meanwhile, AI agents, which have recently become a hot topic in the industry, are being highlighted as a key theme at this year's re:Invent. In the keynote speech the previous day, AWS CEO Matt Garman also stated, "I believe the emergence of AI agents will bring about innovation as significant as the advent of early AI," and emphasized, "AI agents will soon become commonplace in all of our work and daily lives."