About this transcript: This is a full AI-generated transcript of AWS re:Invent 2025 - Keynote with Dr. Swami Sivasubramanian from AWS Events, published June 11, 2026. The transcript contains 13,801 words with timestamps and was generated using Whisper AI.
"The human story is built by those compelled to create. Those willing to be misunderstood. The ones who defy the limits. What if visionaries had a technology that could do more than respond? What if it could make decisions? And deliver beyond imagination? What if it could give scientists the power..."
[00:00:00] The human story is built by those compelled to create.
[00:00:14] Those willing to be misunderstood.
[00:00:17] The ones who defy the limits.
[00:00:20] What if visionaries had a technology that could do more than respond?
[00:00:27] What if it could make decisions?
[00:00:32] And deliver beyond imagination?
[00:00:36] What if it could give scientists the power of a thousand labs?
[00:00:43] Or grow an entire workforce by deepening human collaboration?
[00:00:54] Or equip engineers with the tools to venture beyond the beyond?
[00:01:00] At light speed?
[00:01:03] This is Agentic AI.
[00:01:06] The story is now yours to build.
[00:01:13] What will you reinvent?
[00:01:21] Please welcome the Vice President of Agentic AI at AWS, Dr. Swamy Siva Subramanian.
[00:01:38] All right.
[00:01:39] Hello, everyone.
[00:01:41] And for my kids who are staying back in school just to watch me.
[00:01:46] Hi, Annika.
[00:01:47] Hi, Aruntha.
[00:01:48] But now go back to school.
[00:01:51] All right.
[00:01:53] What an incredible journey Reinvent has been so far.
[00:01:57] I'm excited to share what we have in store for you today.
[00:02:00] But now, let's try something different.
[00:02:04] I want you to close your eyes for a moment.
[00:02:08] I want you to think back to the time you first successfully built your first program.
[00:02:16] Do you remember that sense of accomplishment and freedom?
[00:02:21] That you can create anything you want.
[00:02:24] That exhilarating feeling that you just unlocked a whole new world of possibilities.
[00:02:30] Now, open your eyes.
[00:02:33] For me, it was when I first wrote a program for a scientific calculator.
[00:02:38] Back in the early days of my high school.
[00:02:41] You might be wondering, like, what a bizarre choice.
[00:02:43] Why did you pick that?
[00:02:45] It's because I didn't have a scientific calculator.
[00:02:47] I had a simple kind.
[00:02:49] But then, it took me a few tries before I got it right.
[00:02:53] As I didn't have a computer in my home when I was growing up.
[00:02:57] And my high school had one computer that all of us shared.
[00:03:00] So, I got access to it like 10 to 20 minutes a week.
[00:03:04] So, I had to type first whenever I got access to it and then try.
[00:03:09] And of course, I didn't get it right the first time.
[00:03:11] Because I didn't get all the corner cases of string parsing and various things right.
[00:03:16] But then, when I finally got it right, I was so proud that I built something I didn't have before.
[00:03:23] Suddenly, the kid in me felt that I could solve extremely hard problems
[00:03:29] and build things I didn't have.
[00:03:32] Finally, I was free to imagine and create.
[00:03:36] Today, builders around the world are experiencing that same sense of freedom with the help of AI agents.
[00:03:44] But what makes this moment special and a game changer in the tech industry, in the history of tech, are two things.
[00:03:54] First, who can build is rapidly changing.
[00:03:58] We are no longer constrained by the familiarity of the syntax of language or having to remember hundreds of API calls and parameters and whatnot.
[00:04:09] Second, how quickly you can build is also changing.
[00:04:13] What used to take years now are months.
[00:04:17] What used to take months are now weeks, if not days.
[00:04:20] We are living in times of great change.
[00:04:22] For the first time in history, we can describe what we want to accomplish in natural language.
[00:04:29] And agents generate the plan.
[00:04:31] They write the code, call the necessary tools, and execute the complete solution.
[00:04:36] Agents give you the freedom to build without limits, accelerating how quickly you can go from idea to impact in a big way.
[00:04:47] Take, for example, ocean cleanup.
[00:04:50] That's tackling a really important planet-wide problem by removing plastic debris from our oceans and rivers.
[00:04:58] They are harnessing the power of AI to protect marine life and ecosystems, while also preventing micro-plastics from entering our food chain.
[00:05:08] Their AI system is able to optimize plastic detection models, predict debris movement patterns, and maximize the efficiency of their cleanup operations so that they can operate in the most impactful locations around the globe.
[00:05:24] Or the Allen Institute, that's developing an advanced neural net model for analyzing single-cell multi-modal brain cell data in hopes to unlock groundbreaking new therapies.
[00:05:38] But the reality is, building and scaling these amazing agentic systems are harder than the problems they are trying to solve.
[00:05:48] So let's start by looking at the tools, protocols, and frameworks that make it easy for you to build reliable, accurate, and scalable agentic systems end-to-end.
[00:06:01] But first, let's take a look at what is an agent.
[00:06:05] Agents are those that can sense and interact with a digital environment, and they can convert high-level objectives into a series of executable steps.
[00:06:16] And they constantly learn and improve their efficiency over time.
[00:06:21] But many of you might be wondering, how is an agent different from a generative AI chatbot?
[00:06:27] Now, imagine that over the course of the last few days, the traffic to your website decreased by 40%.
[00:06:34] Now, if you were to ask a chatbot saying, like, why did my traffic go down?
[00:06:38] It would respond to something similar to this, saying, like, that's really concerning.
[00:06:42] You should check your analytics to see which pages were affected.
[00:06:45] And, of course, review any recent changes and look at your server logs and so forth.
[00:06:51] Really nice, helpful advice, but not at all useful.
[00:06:55] But then, agent, on the other hand, will pull analytics data to identify affected pages.
[00:07:01] They query your deployment system for recent code changes.
[00:07:05] They scan the server error logs, and then once it identifies the issues, it will create a bug to get with the problem and the affected code, and even provide a fix for the engineering team to review and deploy.
[00:07:18] In effect, the chatbot tells you what to investigate, whereas the agent investigates, diagnoses the problem, and initiates the solution.
[00:07:28] Today, agents are transforming industries, from software development to drug research, from precision agriculture to architectural design.
[00:07:39] The ability to use and manipulate interfaces, the same way we as humans do, is fundamentally changing how we build applications, dramatically lowering the technical barriers to creating powerful, useful solutions.
[00:07:56] Now, at their core, every agent is built from three important components.
[00:08:03] The first is a model that serves as the brain of an agent, responsible for reasoning, planning, and execution.
[00:08:12] Second, the code that identifies or defines the agent's identity.
[00:08:18] It establishes its capabilities and guides through the decision-making process.
[00:08:24] And then the third is the tools.
[00:08:26] These are the things that bring agents to life, whether they are the backend APIs or access to knowledge base and databases, code interpreters or web browsers that enable real-world action.
[00:08:39] The true power lies in how you orchestrate all these components together.
[00:08:45] Historically, wiring these AI components together was a really labor-intensive and brittle process.
[00:08:53] Yes, frameworks helped, but they still demanded deep expertise.
[00:08:58] Developers found themselves hard-coding complex decision trees.
[00:09:03] These are the intricate state machines with rigid, predefined workflows for every conceivable scenario.
[00:09:10] This created mountains of boilerplate code that was not only difficult to maintain, but nearly impossible to adapt in a dynamic world.
[00:09:20] Because models couldn't reliably reason about the next step, developers had to manually orchestrate every single tool call and state transition.
[00:09:31] And when something unexpected happened, and it always did, agents would often fail.
[00:09:37] But then as models became more sophisticated and gained true reasoning capabilities, everything changed.
[00:09:45] When errors occurred or there was ambiguity, the model didn't just fail.
[00:09:50] It could reason through the alternatives.
[00:09:53] We didn't need static orchestration anymore.
[00:09:56] But in this new world, we believed how you build agents should be really simple.
[00:10:01] Where the agent autonomously plans its actions, they chain-tool all the tools creatively, and they adapt in real-time to unexpected situations.
[00:10:12] And the developer simply defines these components, the model, the code, and the tools.
[00:10:18] This is what inspired our teams to build the strands agent SDK.
[00:10:23] As our teams started building agents for our own products, we took a step back to rethink what the agent SDK of the future would look like.
[00:10:33] One that enables developers to create agents with minimal code.
[00:10:38] Strands embraces a model-driven approach where modern LLMs can autonomously handle any scenario an agent might face.
[00:10:47] This eliminates the need for all these predefined workflows and complex orchestration code, while also improving agent accuracy and code maintenance.
[00:10:57] And boy, strands really delivered.
[00:11:00] We eliminated thousands of lines of code across our agent SDK systems, while improving agent accuracy and code maintenance.
[00:11:09] Realizing its potential, we opened source strands so we could empower every developer to build their own agents.
[00:11:18] And since launching in preview in May, we have worked tirelessly with the community to add new features and capabilities.
[00:11:27] We have hundreds of community contributions to help add features, such as new model providers, A2A support, multi-agent hooks, so you can build more robust and extensible agents.
[00:11:40] And this week, today, we have added two new exciting capabilities.
[00:11:46] The first is support for TypeScript, extending strands to one of the world's most popular programming language.
[00:11:52] And the second is support for Edge devices.
[00:11:55] And the second is support for Edge devices.
[00:11:59] You should check out the Cool Robotics Devon, where they are unlocking new agent capabilities in automotive, gaming, and robotics, enabling autonomous AI agents to run at the edge.
[00:12:12] Developers across every industry now are loving the simplicity and extensibility of strands agent.
[00:12:20] So it's no surprise that in just a period of very few months, strands has been downloaded more than 5 million times.
[00:12:29] And today, we live in a world where almost every developer is starting to experiment and build AI agents on their laptops.
[00:12:40] With all so much activity going on, and so many proof-of-concepts are happening, now leaders are asking why, when they can see all these amazing agents running in production.
[00:12:55] The problem is that most experiments and proof-of-concepts are not designed to be production-ready.
[00:13:02] We need to close this gap and break out of this proof-of-concept jail.
[00:13:07] The first thing you need to get to production faster is that you want the ability to rapidly deploy agents at scale.
[00:13:15] This means having the infrastructure that can rapidly go from zero to thousands of concurrent sessions.
[00:13:23] And it's the ability to support all these long-running agents with the right session isolation that can stop sensitive data from being shared with other agents.
[00:13:33] Next, your agents must navigate massive amount of data and edge cases.
[00:13:39] To get to production, your agents need sophisticated memory systems that can manage context within conversation and interactions, and remember user preferences across sessions.
[00:13:52] Then comes the challenge of identity and access management.
[00:13:57] Without proper security controls, agents can inadvertently access or expose sensitive data that they shouldn't.
[00:14:05] In production, you need rock-solid identity and access management to authenticate users and authorize which tools agents can access on their behalf and manage these credentials across AWS and third-party services.
[00:14:21] And then, of course, as you move to production, your agent is not going to live in isolation.
[00:14:26] It'll be part of a wider system, one that can't fall apart if an integration breaks.
[00:14:33] So you need seamless tool connectivity that enables agents to securely integrate with APIs, databases, and services that your application needs.
[00:14:44] And finally, you need the ability to observe and quickly debug if issues arise.
[00:14:50] Because you simply can't fix what you can't see.
[00:14:55] POCs are never designed with these requirements in mind.
[00:14:59] And most of the shell solutions are not equipped to handle all these changes effectively.
[00:15:04] They are rigid and they lack modularity.
[00:15:07] So you end up building custom logic, duct taping all these solutions together, adding unnecessary complexity that turns your really cool prototype into a maintenance nightmare
[00:15:19] that slows down the innovation.
[00:15:22] That's why we built Amazon Bedrock Agent Code.
[00:15:26] It's the most advanced agentic platform to build, deploy, and operate agents securely at scale.
[00:15:32] It's like having a toolbox where every tool is built for the problems that you will deal with while building and operating agents.
[00:15:41] And people love it because it works with any agentic framework and any model, giving you the freedom to use the tools that work best for your use case.
[00:15:52] And it's modular so that you can mix and match using just the pieces you need for your solution.
[00:15:58] And Agent Code does the heavy lifting.
[00:16:01] So you can focus on what matters most, creating these breakthrough experiences that solve real-world business problem.
[00:16:11] Now, let me show you the difference this makes in practice.
[00:16:15] Take the example of identity and access management, one of the notorious time sinks that can derail any project.
[00:16:22] With Agent Code Identity, you get seamless identity access management across AWS apps and third-party apps like Slack and Zoom, all in just a few lines of code.
[00:16:34] Now, imagine building the same capability from scratch.
[00:16:39] Picture weeks of authentication flows, security protocols, API integration, endless edge cases.
[00:16:46] Which one would you rather write and maintain?
[00:16:49] The ease of agent building that comes from using Agent Code and strands is already helping customers transform their business, like Cox Automotive.
[00:16:59] They have reinvented how they build and deploy agents across their entire organization.
[00:17:05] Now, let's see their story.
[00:17:09] If you think about running a dealership or an auto manufacturer, those have back-office processes.
[00:17:15] Oftentimes, those are manual, they're cumbersome.
[00:17:18] At Cox Automotive, we are in hyperdrive, moving aggressively towards disruptive, positive value to change the game, change how you shop for a car, change how you sell a car.
[00:17:29] Change how you service a car, and helping our customers level up their own operations in ways they never thought possible.
[00:17:36] We also took an internal use case with an agentic solution that we call Fleetmate.
[00:17:41] So, something that actually took two days per estimator, we got that down to less than 30 minutes per vehicle.
[00:17:48] AWS really shows up as a true partner.
[00:17:51] Agent Core has allowed us to be able to move fast, building agents, allowing agents to work with other agents.
[00:17:57] They are allowing us to be inside the product roadmap and help shape what we need in order to innovate at the speed that our customers are expecting us to.
[00:18:06] Bringing agents to life allowed us to work across our different areas of business and find opportunities where we could bring agentic thinking into a workflow and drive value.
[00:18:16] One of the most exciting things about our agentic story is seeing an engineer using tools like Agent Core, Strands, and Bedrock and building agentic solutions.
[00:18:26] Their mind gets blown.
[00:18:28] They can go from not knowing the tool, not knowing how to use it, and producing something in days, sometimes hours, things they never thought were possible.
[00:18:37] And that is super exciting to see.
[00:18:42] Wow.
[00:18:44] It's exciting to see how Cox transformation is opening a world of possibilities.
[00:18:49] It's stories like these that fuel our innovation engine.
[00:18:53] And just yesterday, Matt unveiled two new capabilities that came directly from you, in terms of feedback.
[00:19:00] Agent Core policy, that provides control over agent interactions and behavior, while giving agents the freedom to reason and take the best action to fulfill a request so that you can build trusted agentic capabilities.
[00:19:15] And agent core evolves, which enables evaluating and testing agents in thousands of simulated scenarios before they meet your customers.
[00:19:26] So no longer do you have to cross your fingers and hope for the best when you launch an agent.
[00:19:31] But today, I want to explore something that gets to the heart of what makes agents truly intelligent: memory.
[00:19:39] Think about the experience at your favorite local restaurant.
[00:19:45] What made it exceptional?
[00:19:47] It's because they remembered you.
[00:19:50] They know your name.
[00:19:51] They know your seating preferences.
[00:19:54] And they know your favorite meal.
[00:19:56] In the agentic world, we have short-term memory handling, the immediate conversational flow, and long-term memory capturing the insights across sessions.
[00:20:06] But something was missing: the when and the why behind user behavior.
[00:20:12] For example, picture this: you're rushing to catch a flight for work.
[00:20:17] As a solo traveler, I travel carry-on only, and I like to be the last one on the plane.
[00:20:23] That means my agentic travel agent books my right to arrive 45 minutes before departure with all my TSA free and clear and whatever else will come.
[00:20:34] I will do it with perfect timing.
[00:20:37] But then, pass over three months.
[00:20:39] But this time, I'm wrangling a four-year-old and a ten-year-old, plus enough luggage with my wife for a small army.
[00:20:47] My agent needs to remember the chaotic family trip from last summer.
[00:20:53] It needs to recognize the pattern and automatically book my ride to arrive at the airport at least two hours early.
[00:21:01] Because traveling with kids changes everything.
[00:21:04] So what you need is not just the memory of the past, but also the context of the current interaction.
[00:21:12] That's why we launched Episodic Memory, a new functionality with Agent Core long-term memory so that agents can remember and learn from the past experiences.
[00:21:24] Thank you.
[00:21:30] This new feature gives the ability to build agents that truly understand user behavior, and they adapt automatically by recognizing patterns across similar situations, and they proactively offer solutions that work.
[00:21:46] It enables agents to store and recall specific experiences or interactions as discrete episodes, similar to how we remember particular events in our lives.
[00:21:59] The more your agent's experience, the smarter they become.
[00:22:03] Now, let's look at how customers are unlocking the freedom to innovate with Agent Core.
[00:22:08] Heroku, who helps companies build and deploy apps in the cloud with their AI as a platform service, they use Agent Core to build Heroku Wives, enabling anybody to build and run entire web apps using natural language.
[00:22:24] With Agent Core run time, they were able to speed up their development velocity and went from proof of concept to launch in just five weeks.
[00:22:34] And the PGA2, a pioneer and innovation leader in sports, they have built a multi-agent content generation system to create articles for their digital platform.
[00:22:46] Their new solution built on Agent Core enables the PGA2 to provide comprehensive coverage for every player in the field by increasing content writing speed by 1000%, while also achieving a 95% reduction in cost.
[00:23:03] And finally, Kalent, a next-gen cloud services company that helps companies build and scale their cloud journey.
[00:23:11] By rebuilding on Agent Core, Kalent deleted thousands of lines of custom orchestration code, routing and memory, and replaced it with a managed serverless agent platform that just works.
[00:23:24] Now they are shipping new integrations in days instead of weeks.
[00:23:29] They will cut their ops overhead by 70% and unlock new capabilities like 8-hour workflows and multi-turn agent reasoning that simply weren't possible on their custom-built system.
[00:23:41] Next, I would like to invite William Brennan, VP of Tech Transformation at Blue Origin, to explain how they are leveraging AWS for their AI-powered space exploration.
[00:23:53] Good morning, everybody.
[00:24:04] At Blue Origin, we're building the road to space for the benefit of Earth.
[00:24:09] In pop culture, space explorers have always had AI partners.
[00:24:14] Just think of some of those famous explorer duos from your childhood.
[00:24:18] The farm boy and his droids, the starship captain and his voice-activated ship's computer, and yes, even the mission commander whose AI went rogue.
[00:24:27] We've long dreamed of a world where artificial intelligence and space exploration go together.
[00:24:33] From rockets to AI, at Blue Origin, we're not dreaming of the future anymore.
[00:24:37] We're building it.
[00:24:39] The new Shepard rocket became the first spaceship to autonomously launch and land vertically, making automation and reusability safe and reliable partners.
[00:24:49] Since then, 86 humans have trusted their lives to New Shepard's automation across 15 flights, experiencing a view of Earth they've always dreamed of.
[00:24:58] The road to space grew visibly and powerfully in last month's successful launch and landing of our new Glenn orbital rocket.
[00:25:07] This next frontier, an increased launch cadence, lunar landers, a permanent presence on the Moon, is building on that automation and extending it using the latest innovations in AI.
[00:25:21] Our mission to reduce the cost of access to space is accelerated through agentic AI that learns, adapts, makes decisions, that reasons, plans, and explains its thinking.
[00:25:32] That's why, frankly, I think I literally have the coolest job in the world.
[00:25:36] We're going back to the Moon, this time to stay.
[00:25:39] And agentic AI is one of our new teammates helping us get there faster and cheaper.
[00:25:44] Across Blue, the road to space is being built by incredibly talented and creative people, engineering, manufacturing, software, supply chain.
[00:25:53] As a vertically integrated company, we run the full life cycle of a space mission, from concept to operations.
[00:25:59] Building and launching rockets requires tremendous specialization from dozens of these disciplines, and that practical knowledge isn't found in general large language models.
[00:26:08] It's buried in our systems, databases, and tribal knowledge.
[00:26:12] Agendic AI's ability to create teams of domain expert agents with access to specific knowledge and custom tools has completely transformed the application of AI at Blue.
[00:26:23] Taking a democratized approach to AI, everyone at Blue is expected to build and collaborate with AI agents to do their work better and faster.
[00:26:32] By equipping our teams with knowledgeable and capable agents, we are able to dramatically accelerate the product lifecycle, increase production rate, and most importantly, reduce the cost of access to space.
[00:26:45] The starting point for our builders and creators is our internal generative AI platform we call Blue GPT.
[00:26:51] It's a secure... There we go, Blue GPT team.
[00:26:54] It's a secure LLM and MCP gateway, agent marketplace, and multi-agent orchestration system built on core AWS primitives.
[00:27:02] Bedrock, strands agents, EKS, OpenSearch, RDS, Lambda, giving our agents access to all leading models, custom knowledge bases, and dozens of tools.
[00:27:12] Agendic AI has exploded at Blue through organic, disruptive adoption.
[00:27:18] Over 2,700 agents have been created and are in use and production in the agent marketplace, driving over 3.5 million interactions across 70% of our employees in the last month.
[00:27:30] There are so many examples I could give on how Team Blue is using agents to work better and faster.
[00:27:35] 95% of our software engineers are using agents to write code.
[00:27:39] New Glenn is using agents to accelerate launch approvals.
[00:27:43] Supply chain and manufacturing use agents for design changes and work orders.
[00:27:48] But what I really want to talk about are T-Rexes.
[00:27:51] No, not space dinosaurs, but rather an innovative project of Blue Origin.
[00:27:56] T-Rex stands for Thermal Energy Advanced Regolith Extraction.
[00:28:01] And it's a brand new capability being developed by our in-space exploration systems group to solve one of the Moon's toughest challenges,
[00:28:08] surviving the lunar night, equivalent to 14 Earth days of freezing darkness.
[00:28:13] I have part of it right here with me, and this thing is amazing.
[00:28:17] As you can see in the video, the battery, the T-Rex, takes lunar regolith or Moon dust from the surface of the Moon,
[00:28:25] circulates it through this chamber, and then extracting the heat through a lightweight heat exchanger,
[00:28:30] and then runs it through this cylinder to save the rest of the machine from the sensitive and abrasive rocks.
[00:28:38] Designed to run for years, this heat cycle is then reversed during the lunar day to recharge regolith for use during the next night,
[00:28:45] turning Moon dust into a battery.
[00:28:49] While T-Rex is amazing, what's even more amazing is the team of AI agents that helped build it.
[00:28:55] A Blue GPT agent helped us with the detailed requirements.
[00:28:59] Another agent helped us create the system architecture.
[00:29:01] One of my favorite agents, though, is our simulation and analysis agent,
[00:29:05] where we're able to automate the iterative design loops, rapidly improving the product.
[00:29:10] With custom knowledge bases and connected to AI-native tools like NTOP for mass and performance optimization,
[00:29:17] and Astari Digital for deterministic validation and artifact management,
[00:29:21] the agents run complex physics simulations, iterate on the design based on the results,
[00:29:26] and execute until the requirements are met, all running on GPU-accelerated EC2.
[00:29:31] T-Rex is a great example of what we see as the future of engineering teams at Blue.
[00:29:37] Small teams of experts working with large teams of AI agents to deliver the works of 10s,
[00:29:42] and at orders of magnitude faster than before.
[00:29:45] The benefits are immense.
[00:29:47] Using agents, our engineers were able to deliver a performant product over 75% faster than traditional approaches,
[00:29:54] with mass 40% improved over the original designs.
[00:29:59] Now, I get it. Most of you aren't building rockets or lunar infrastructure, but you do deal with specialized knowledge and custom tools,
[00:30:05] and I think could benefit from a few of our learnings with agents.
[00:30:08] First, I'd consider you to think of adoption of AI as everyone's job, not just the technology team's.
[00:30:15] Democratize agent creation and use.
[00:30:18] A simple, user-friendly interface led to 70% company adoption.
[00:30:21] Second, I encourage you to let the process for innovation be messy, but discovery easy.
[00:30:27] Third, a public marketplace or agent app store has really helped us here, providing easy-to-use eval tools to allow your teams to test and discover their agents.
[00:30:36] Third, let AWS handle the LLM, agent primitives and scaling, and focus on building with those tools to solve problems for yourselves and for your customers.
[00:30:47] With Agentic AI, we believe that the future is now, it's just not evenly distributed.
[00:30:53] We aren't yet developing all of our products using AI, but we will be.
[00:30:57] With millions of people living and working in space, we believe in a world where we can agentically design an entire rocket, where we can launch 100 rockets with one person, rather than one rocket with 100 people.
[00:31:10] Our work with AWS isn't slowing down.
[00:31:13] We are building an enterprise knowledge graph on AgentCore, transforming the AI accessibility of all of our data, and we're excited to test some of our own custom-built models with our autonomous lunar rover on the moon, trained on AWS GPUs.
[00:31:26] Our talented team of Blue Origin employees and AI agents will continue to be the fuel for this rapid innovation as we build the road to space for the benefit of Earth.
[00:31:37] Thank you very much.
[00:31:39] Thank you, William.
[00:31:41] It's fascinating to see how customers are leveraging agents, even in space exploration.
[00:31:59] Now, as agents become easier to build, the next big question emerges: How do we make them more efficient?
[00:32:07] Today's off-the-shelf models have broad intelligence.
[00:32:10] They can handle complex tool use, multi-step reasoning, and unexpected situation.
[00:32:16] But they aren't always the most efficient.
[00:32:18] And this efficiency is not just about cost.
[00:32:21] It's about latency: How quickly can your agent respond?
[00:32:25] It's about scale: Can it handle peak demand?
[00:32:28] It's about agility: Can you iterate and improve quickly?
[00:32:31] And these are the practical realities of deploying AI at scale.
[00:32:36] So, how do we address these challenges?
[00:32:39] The majority of the agents, depending, they spend most of the time on routine operations.
[00:32:45] These are like writing code, analyzing search results, creating content, and executing predefined workflows.
[00:32:52] If we know these stars up front, we can customize models specifically for them.
[00:32:57] Model customization enables us to build agents that are efficient to deploy at scale while achieving the right performance.
[00:33:06] So, the question isn't whether you should customize your models, but how quickly can you get started.
[00:33:13] So now, let's explore the techniques available and how we can help you turn this vision into reality.
[00:33:21] The first technique we will look at is supervised fine-tuning.
[00:33:25] Think of it as teaching your AI agent to be a specialist instead of a generalist.
[00:33:30] Like turning a family doctor into a cardiologist who is laser-focused on exactly what you need.
[00:33:37] This is where most teams start, and for really good reason.
[00:33:40] It's practical.
[00:33:42] You don't need massive compute budgets or a PhD in machine learning to get started.
[00:33:47] You start with a pre-trained model and train it on curated, agent-specific data sets containing tool use pattern, multi-step reasoning traces, and successful task completion.
[00:33:59] This creates permanent behavior changes, and they don't require lengthy prompts and can dramatically improve performance on specific tasks where the base model struggles.
[00:34:12] But here is the catch.
[00:34:14] Your model can get so focused on specific tasks due to overtraining that it might lose those amazing capabilities that made you choose the big model in the first place.
[00:34:25] So the key to success here is quality over quantity.
[00:34:30] A data set with 10,000 carefully curated agent interactions will outperform millions of generic examples.
[00:34:39] Think of it this way.
[00:34:41] Would you rather learn surgery from 10,000 recordings of world-class surgeons performing your exact procedure or from millions of random YouTube videos?
[00:34:51] The choice is obvious.
[00:34:53] The next technique is model distillation.
[00:34:57] You should use model distillation when you have deployment constraints like memory limitation and you need to deploy a smaller, faster model.
[00:35:06] Think of this method as something like creating a brilliant apprentice who learns from a master craftsman, but they can work twice as fast with half the tools.
[00:35:17] So you take a large, powerful teacher model and train a much smaller student model to mimic not just the answer, but also how they think through and reason.
[00:35:27] The student learns into their confidence levels, decision patterns, and even recognition strategies.
[00:35:33] Here, we often see up to 10x speed up with models retraining 95 to 98% of the teacher's performance, and that's the difference between a three-second latency for customer calls versus getting an almost instant response.
[00:35:51] The trade-off, you need that high-performing teacher model, and distillation requires substantial compute resources.
[00:36:00] Once you get that efficient distilled model, it pays dividends with every deployment.
[00:36:06] And then you have reinforcement learning, where models learn from the outcomes of their actions by getting rewards for good outcomes and penalties for bad ones.
[00:36:18] And there are two flavors of RL, RL from human feedback and RL from AI feedback.
[00:36:24] With human feedback, here you show people multiple AI responses to the same customer inquiry and ask them to rank from the best to worst.
[00:36:35] Now we use these rankings to train a reward model that learns to think like these human evaluators, becoming your automated quality scorer.
[00:36:45] Here is how it works. Your agent executes an action, and then the reward model gives it a gold star and says, "Too bad, try again."
[00:36:54] The agent learns through trial and error, just like we do in many cases.
[00:36:59] This approach shines when outcomes are impossible to define programmatically.
[00:37:04] How do you code up a sound empathetic, but professional? You simply can't.
[00:37:09] But humans know it when they see it, and now your model can learn to recognize it too.
[00:37:14] In reinforcement learning with AI feedback, instead of human rankings, we use a powerful LLM to evaluate and rank these responses.
[00:37:25] This is much faster, much cheaper, and more consistent than humans.
[00:37:29] It's perfect for outcomes with clear right or wrong answers, and lets us scale to millions of examples.
[00:37:36] It also rewards good processes, not just the answers.
[00:37:41] The model learns to think strategically, step by step, which is exactly what agents need.
[00:37:46] Think about it. When an agent is troubleshooting a complex issue, you don't just want the final answer.
[00:37:53] You want the logical diagnostic steps, the right clarifying questions, and methodical problem solving.
[00:37:59] That is strategic thinking, and that's what RL teaches.
[00:38:04] While RL is incredibly powerful, implementation of RL techniques can be a challenge.
[00:38:11] You need PhD-level expertise in reward modeling, policy optimization, and integrating feedback.
[00:38:18] Then there is setting up complex distributed infrastructure, and spending months in development with no guarantee it is ever going to work.
[00:38:27] And the cost.
[00:38:29] We will be talking six to 12 month projects running on high-end compute resources, and most companies simply can't afford it.
[00:38:38] So you have to compromise.
[00:38:40] You either settle for the generic models with poor performance, or you spend millions customizing models with reinforcement learning.
[00:38:48] And then when your workload patterns change, you start this entire expensive process all over again.
[00:38:55] So we thought, what if we could take away all the complexity and cost while still giving you access to these advanced training techniques?
[00:39:05] That's why today we are making reinforcement fine tuning available in Amazon Bedrock.
[00:39:15] Reinforcement fine tuning RFT helps customers improve model accuracy without needing deep machine learning expertise or large sense of labeled data.
[00:39:26] Bedrock automates the entire RFT workflow by making this advanced model customization technique accessible to everyday developers.
[00:39:36] Here, you can start today with Amazon Nova, and we will be starting support for open model weights really soon.
[00:39:45] Here, it works very simple.
[00:39:46] You start by selecting the base model you want to modify, point it at your Bedrock logs, and select a reward function.
[00:39:54] The easiest of which is an LLM-based charge.
[00:39:57] And that's it.
[00:39:59] Behind the scenes, Bedrock handles all the complexity of reinforcement learning implementation,
[00:40:05] it's striking the right balance of customization and ease of use for most use cases.
[00:40:12] And it provides delivering 66% accuracy gains on average over the base models.
[00:40:19] That is how powerful these techniques are.
[00:40:22] Now, while Bedrock offers an easy and powerful way to fine tune models,
[00:40:28] there are many more customers that want the full control over their customization techniques and want to leverage that data.
[00:40:35] Think about it.
[00:40:36] A legal firm's competitive advantage is its accumulated knowledge base and reasoning patterns.
[00:40:43] Here, training a model from scratch ensures this proprietary expertise remains confidential, while also giving them a competitive advantage.
[00:40:51] A healthcare provider needs models trained on the specific patient outcomes, not just generic medical data.
[00:40:59] That's exactly why we built SageMaker AI.
[00:41:03] SageMaker gives you everything you need to build, train, and deploy AI models that are uniquely yours at whatever scale.
[00:41:12] It supports knowledge distillation, supervised fine tuning, and direct preference optimization.
[00:41:18] You can do full weight training or parameter efficient fine tuning, whatever your use case demands.
[00:41:25] With SageMaker's IDE and MLOps capabilities, you can go from idea to production in weeks, not months.
[00:41:33] Now, customers like Dhan India, who built a large language model that deeply understands the complexities of Indian financial market.
[00:41:43] They started with Mistral 7B as a base model.
[00:41:47] They built ArthaM, a specialized 7B parameter model, using a combination of training methods like continual pre-training, fine tuning, and knowledge distillation.
[00:41:57] Their solution used Amazon SageMaker model.
[00:41:58] Their solution used Amazon SageMaker for the model building and training, and used Bedrock for the teacher model support.
[00:42:06] The new customized model runs on a single GPU infra and outperforms state-of-the-art models 88% of the time, while providing better performance at a fraction of the operational cost.
[00:42:20] However, the very flexibility that SageMaker AI makes it so capable also means that you must navigate numerous decisions and technical hurdles to achieve your goals.
[00:42:34] Think about what you are actually signing up for when you want to customize a model with your data.
[00:42:40] First, you are defining the goals and picking evaluation criteria.
[00:42:44] Are you optimizing for accuracy or latency or cost?
[00:42:48] Then you are choosing fine-tuning, rag, or maybe pre-training from scratch.
[00:42:53] Then comes the fun part, data prep.
[00:42:56] You are cleaning data sets, checking for biases, making sure everything is properly formatted.
[00:43:01] And anyone who has done this, it's never as clean as you hope it will be.
[00:43:06] And then you are spinning up infrastructure, configuring training jobs, watching these loss goals, and tweaking all these hyperparameters.
[00:43:14] And when something breaks at 2:00 AM, guess who is getting that alert?
[00:43:19] Finally, after weeks of work, you have a model.
[00:43:23] But wait.
[00:43:24] Now you need to evaluate it and optimize it for production and making sure it meets your company's AI governance standards.
[00:43:32] Then the result, what should be a really exciting AI project, turns into months of laborious, challenging work before you see any business value.
[00:43:42] What if we could compress that journey from months down to just days?
[00:43:48] That's why today, I'm thrilled to announce the release of new serverless model customization capabilities in SageMaker AI.
[00:43:57] With this release, you can customize popular models such as Amazon Nova, Quen, Lama, DeepSeq, and deploy them directly on Bedrock or SageMaker in just a few steps.
[00:44:14] BetterEd, this comes with two experiences.
[00:44:18] And you can choose the right approach based on your comfort level.
[00:44:21] A self-guided approach for those who like to be in the driver's seat.
[00:44:25] And an agent-driven experience that uses an AI expert for folks who like to turn on autopilot in their cars.
[00:44:33] Each of these parts provide you access to new RL techniques like RL AI/FF, RL VR, and DPO, and are based on the best model for the job, either your choice or the agents.
[00:44:47] Now, the agent AI experience launching in preview removes the heavy lifting so that you can focus on the outcome.
[00:44:55] You use natural language to explain your use case, and the AI agent guides you through the full customization workflow.
[00:45:03] It first analyzes your scenario to recommend the right fine tuning technique.
[00:45:08] Then if needed, it will generate the synthetic dataset for your model customization.
[00:45:13] And then it sets up the entire serverless infra to train the model without any manual intervention.
[00:45:20] And then finally, it evaluates the trained model against the base model to determine if the customization was successful.
[00:45:28] So now, what used to take, like many ML engineers and months of trial and error, now happens in just days.
[00:45:38] All guided by an agent that knows the best practices and does the heavy lifting for you.
[00:45:45] Now, while these model customization techniques enables you to build agents that are efficient to deploy at scale,
[00:45:53] working closely with customers helped us identify another key gap.
[00:45:59] Traditional techniques can't deliver the domain-specific intelligence, accuracy, and price performance that competitive applications need.
[00:46:09] Especially when you need models that deeply understand your industry, data, and workflows.
[00:46:16] For example, a drug discovery company needs an AI model that deeply understands molecular structures.
[00:46:23] Protein interactions and clinical data specific to their therapeutic area.
[00:46:28] Generic models, they lack their specialized knowledge.
[00:46:32] And fine tuning simply can't embed the foundational understanding of their proprietary research.
[00:46:38] Now, many of you might are saying like, Swamy, why can't you use any of the customization technique we talked about earlier?
[00:46:46] The simple answer is that these techniques add knowledge on top of a model that wasn't designed for that domain.
[00:46:53] It's the difference between teaching a general translator some medical terms and training a medical translator from the beginning.
[00:47:02] This means today, if you want to truly customize foundational models, you end up starting from scratch.
[00:47:09] You need millions of dollars in computes, months of training time, and specialized ML experts.
[00:47:15] You need to have to manage massive infra, and you're starting from zero with no proven baseline or a guarantee of success.
[00:47:22] This is why custom foundational models have been the exclusive domain of well-funded organizations and AI startups.
[00:47:30] But now with NOVA Forge, which Matt launched yesterday, you have a first-of-its-kind program that offers you the easiest and most cost-effective way to build your own frontier model with Amazon NOVA.
[00:47:44] You get access to intermediate checkpoints, and you can mix in your proprietary data with Amazon-curated data during mid-training cycle without having to worry about the compute, data, and time needed for the full model training.
[00:48:00] The result, a model with frontier intelligence tailored to your industry and use case while retaining NOVA's foundational knowledge, safety, and reliability, all without the cost and effort of the full training lifecycle.
[00:48:15] Now, another area we are focused on a lot is to help you build and run cost-effective and performance agencies' infrastructure.
[00:48:26] This is where SageMaker HyperPod comes in.
[00:48:30] HyperPod has transformed the traditional complex and trying-consuming process of managing infra for training and deploying models into a fully managed service.
[00:48:41] It removes the undifferentiated heavy lifting and scales across thousands of AI accelerators with automatic workload prioritization and reduces model development costs by up to 40%.
[00:48:54] And most critically, it gives you full visibility and control over how different tasks are prioritized and how compute resources are allocated to each task so that you can maximize these expensive resources.
[00:49:08] However, when you're building or training a model, it's not uncommon to have failures.
[00:49:14] And as your models become bigger and they require even larger training clusters, the failures become even more frequent and they take longer to resolve.
[00:49:25] So, to deal with these unexpected failures, HyperPod automatically saves these model snapshots or checkpoints.
[00:49:32] And when a failure occurs, they recover from a previously saved checkpoint.
[00:49:37] But this traditional checkpoint-based recovery is not always easy.
[00:49:40] You have to pause the entire cluster and then diagnose the issue and then restore from the saved checkpoints.
[00:49:48] All the while, leaving expensive AI resources idle for hours, keeping your CFO up at night.
[00:49:56] So, we thought, what if we could find a way to improve training resiliency and reduce ground time?
[00:50:03] That's why today, we are announcing checkpoint-less training on SageMaker HyperPod.
[00:50:15] With this new feature, we significantly reduced the recovery overhead since the process no longer requires rolling back a partial training, no more than a partial training step.
[00:50:26] This is a paradigm shift in model training that automatically recovers from infra faults in minutes with zero manual intervention, even with clusters that span hundreds of thousands of AI accelerators.
[00:50:40] The way we achieve this is by continuously preserving the model state across the distributed cluster.
[00:50:47] So, when something goes wrong, there is no need to panic and roll back to some old checkpoint.
[00:50:53] Instead, it smoothly swaps out the faulty hardware and uses peer-to-peer transfer to grab the exact model state from healthy accelerators nearby.
[00:51:03] So, now, you get faster recovery and significant cost savings in your model training.
[00:51:10] And now, with all of these launches we have announced today, we are excited to give you the tools you need to build highly efficient models.
[00:51:21] Now, let's hear from Guillermo Rauch, CEO and founder of Vercel,
[00:51:26] who has put these principles into practice, leveraging AWS to build efficient and scalable tools that deliver real value to millions of developers today.
[00:51:44] Hi, everyone.
[00:51:46] It's so great to be here with you all in the cloud.
[00:51:50] When I was growing up in Argentina, I deconstruct sites on our family computer and bring my own ideas to life by FTPing a file to a server and sharing a link with my friends.
[00:52:03] It was the best feeling I could just build.
[00:52:07] And I never stopped building.
[00:52:09] So, I built with every tool and every framework, which eventually led me to create my own framework, as one does.
[00:52:16] Next.js.
[00:52:18] With Next.js, we made building high-quality applications accessible to anyone.
[00:52:24] From the biggest enterprises on the planet to solo developers starting out with just an idea.
[00:52:30] The problem was that even with Next.js, I'd build an app in a couple hours and then spend weeks setting up infrastructure to get it out into the world.
[00:52:41] And I didn't just want easy.
[00:52:43] I wanted peak performance at scale, and I wanted it now.
[00:52:48] That idea led me to build Vercel.
[00:52:52] 10 years later, Vercel now empowers more than 11 million customers to build, scale, and secure global web applications without managing infrastructure.
[00:53:05] With more than 1 trillion requests served every month.
[00:53:10] Now, AI is changing what and how we build.
[00:53:15] Interfaces are becoming conversations.
[00:53:18] And workflows are becoming autonomous.
[00:53:21] We're witnessing the shift from web pages to AI agents.
[00:53:25] This is perhaps the most important transformation we'll see in our careers as developers and technologists.
[00:53:33] And this is how the Vercel AI Cloud was born.
[00:53:37] To redefine fully managed infrastructure for the AI era.
[00:53:43] We call this self-driving infrastructure.
[00:53:47] You write an app, you push your code, and automatically, the best possible infrastructure is provisioned, orchestrated, and optimized for you.
[00:53:58] Vercel operates as an always-on system that observes every input from the code you write, the code you ship, to the traffic your users generate, and turns that data into real-time insights and optimizations.
[00:54:16] You get to focus on your business, your products, your customers.
[00:54:23] And we know this works because we prove it at scale every day with our customers, but also because we build Vercel, guess what?
[00:54:32] With Vercel, on Vercel.
[00:54:35] And Vercel is powered by AWS, which means with just one Git push, you reap the benefits of fully self-driving AWS infrastructure.
[00:54:49] So, let's look at this in action.
[00:54:52] First, V0.
[00:54:54] V0 is the realization of that original dream.
[00:54:58] The thing I wished existed when I was hacking together my websites in Argentina.
[00:55:03] You speak your mind, and in seconds, you have a live, working application.
[00:55:09] Since its launch, V0 has been serving millions of requests per day.
[00:55:15] And that was possible because, under the hood, V0 runs on Vercel, which runs on AWS, with model inference provided by Amazon BadRock.
[00:55:26] When we were building V0, our engineers could just write code, deploy it over cell, and scale without configuring AWS primitives.
[00:55:37] And so, this freed up the team's time, the V0's team's time, to focus on the novel AI challenges, with new models and new ideas and new techniques coming up every week.
[00:55:48] So, to keep up and move at the speed of AI, we built the AI SDK.
[00:55:54] So, think of it as what Next.js did for building pages, AI SDK does for building agents.
[00:56:03] So, first, of course, we built it for ourselves, and then we made it available for everyone.
[00:56:08] It makes AI simple, model-agnostic, and scalable.
[00:56:13] Today, AI SDK is downloaded about 5 million times per week.
[00:56:21] And with AI SDK, anyone can build with AI.
[00:56:25] Whether it's your first AI app, or you're at scale.
[00:56:30] If you're at scale, like Thomson Reuters, their co-counsel AI assistant serves attorneys, accountants, audit teams.
[00:56:39] They used AI SDK, and Vercel, and a team of just three developers build their tax advisory agent in just two months.
[00:56:49] And it's already in use by 1,300 accounting firms.
[00:56:54] Now, they're migrating their entire codebase to AI SDK, deprecating thousands of lines of code across 10 providers, and consolidating everything into one composable and scalable system.
[00:57:10] Three developers, two months, thousands of firms.
[00:57:15] And so, as we move from pages to agents, and pixels to tokens, it's helpful to reuse lessons from the past.
[00:57:26] The web scaled thanks to CDNs, which increased speed and reliability.
[00:57:33] And that's why we built the AI Gateway.
[00:57:36] Think of it as a CDN of tokens.
[00:57:39] It routes inference automatically, it adds failover, and critical controls.
[00:57:45] AI Gateway is built on the same global network that sells all of our customers, across 20 regions, close to your users, and your AWS data.
[00:57:57] This network runs on battle-tested AWS infrastructure.
[00:58:02] Global Accelerator, Shield, S3, EC2, and more.
[00:58:08] But here's the next challenge.
[00:58:11] AI is changing the profile of our workloads.
[00:58:15] While traditional apps need to load in milliseconds, agents can think over minutes, or even hours, and soon days.
[00:58:24] So, serverless has been great for scaling on demand, right?
[00:58:28] So, which we also need for agents.
[00:58:30] But it hasn't been so great for what we call idle time.
[00:58:35] That's why we built Fluid Compute.
[00:58:38] It's compute optimized for AI.
[00:58:42] Powered by AWS, Fluid adapts to every workload, whether it's pages or agents.
[00:58:49] Your functions scale on demand.
[00:58:52] But crucially, you pay only for what you actually use.
[00:58:57] Active CPU time.
[00:58:59] This is fully self-driving compute infrastructure.
[00:59:04] Our AI cloud goes beyond for sales-owned services, providing seamless access to even more AWS solutions, like Agent Core, Hero, and RDS.
[00:59:17] How we build the web is changing.
[00:59:22] You'll be working with agents, like VZero and Kira.
[00:59:26] These agents will be writing code using frameworks, like Next.js and AI SDK.
[00:59:32] And they'll deploy to self-driving infrastructure.
[00:59:36] This is the future of the web, powered by Vercel, backed by AWS.
[00:59:42] Thanks, Guillaume.
[00:59:44] Thanks, Guillaume.
[00:59:46] Vercel is a great example of how quickly agents can help a business grow and thrive.
[00:59:49] Not only are they helping their end users build and run intelligent agents, but they are also helping their builders at Vercel enable entirely new developer ecosystems.
[01:00:01] Now, as we hand over more tasks to AI agents and ask them to act autonomously on our behalf, we need to be able to trust that they will perform things as they expected.
[01:00:22] While most of us have spent our entire careers building an accurate and reliable systems, agents introduce us a whole new set of challenges that require radical new approaches.
[01:00:39] At AWS, one of the first agents we built a couple of years ago was an early prototype for what is now Kira CLI.
[01:00:48] We found that while it worked great, but in some cases, LLNs would hallucinate API calls.
[01:00:55] And of course, we can't just always say, "Hey, it's okay if it hallucinates."
[01:01:00] And then we got together and brought a bunch of our scientists, machine learning scientists, and also we reached out to our automated reasoning team to see if we could solve this problem with what is now we call it as neuro-symbolic AI.
[01:01:15] In effect, we are creating the yang to the LLN, so to speak.
[01:01:20] So, we have a special treat for you today.
[01:01:23] We brought Byron Cook, AWS Distinguished Scientist, and the foremost authority in automated reasoning to explain how we are leveraging automated reasoning across AWS to help our agents be more trustworthy.
[01:01:49] Thank you, Swami.
[01:01:51] All right, so let's start with a question.
[01:01:54] How much do you trust agents today?
[01:01:57] For example, do you trust that they'll send money to the right place?
[01:02:01] Or do you trust that they'll follow local laws when operating on your behalf?
[01:02:08] Giving an agent access to your credit card is like giving a teenager access to your credit card to go use on the internet.
[01:02:16] Like, sure, they'll probably get a lot done on your behalf, but you might end up owning like an agent.
[01:02:21] Like a pony or a warehouse full of candy.
[01:02:26] So, the challenge is that agents are based on large language models, and large language models hallucinate.
[01:02:34] Right? They make mistakes in the face of complex rules or logic.
[01:02:38] Their reasoning contains logical errors.
[01:02:43] So, imagine an agent is interpreting a ticket returns policy, and it's making mistakes.
[01:02:51] Now we're giving away money where we shouldn't.
[01:02:54] That's not the sort of agent we're going to keep in production for very long.
[01:02:58] So, to make matters worse, LLMs can be tricked, right?
[01:03:05] Bad actors can craft inputs to trick LLMs.
[01:03:09] It's actually not hard to do today.
[01:03:11] So, this is because most agentic systems are using statistical methods, like LLM as a judge, to enforce their guidelines.
[01:03:20] That's never going to work in situations where sensitive matters like trust, or money, or human lives are at stake.
[01:03:30] So, when we don't trust agents, we tend to overcompensate.
[01:03:34] Right? We introduce additional human oversight on every step.
[01:03:41] Right? We hard-code the steps that they'll take.
[01:03:47] So, both of these techniques reduce the creativity and the autonomy that agents have,
[01:03:55] which is precisely the reason we're excited about agents in the first place.
[01:04:00] So, what we want is an easy way to specify our constraints on agents at the outset.
[01:04:05] Right? Giving agents as much freedom as possible, while defining the envelope in which they can operate safely.
[01:04:12] We then want assurance that they'll follow those constraints, even when the constraints are subtle or complex.
[01:04:19] And so, the good news is that we have much of the technology we need for this purpose.
[01:04:24] And that's why Swami has asked me to come up on stage today to tell you about my scientific discipline of automated reasoning.
[01:04:30] And how we're applying it to agentic AI.
[01:04:32] So, automated reasoning is the search for and also the very detailed checking of proofs in mathematical logic.
[01:04:40] It's the same type of reasoning, you know, performed by people like Euclid during the time of the ancient Greeks.
[01:04:45] So, in our context, we can reason about all possible executions of a computer program in the same way that Euclid reasoned about all possible right-angle triangles when proving the Pythagorean theorem.
[01:04:57] So, we've come full circle in a sense, right, we're taking ideas from 2000 years ago and applying them to one of today's most vexing challenges in the tech space.
[01:05:08] So, consider a system where we want to know that the data is encrypted before it's stored.
[01:05:14] So, we can encode that property into a logical formula, namely like a temporal logic, and then use techniques from mathematical logic to prove that that property is in fact always true.
[01:05:27] Furthermore, we're using symbolic techniques, symbolic logic, and that allows us to know that we're reasoning about all possible inputs to the system and also all possible configurations that the system might reach during its execution.
[01:05:41] So, within AWS, we've actually been using this technique for over a decade.
[01:05:46] We've been reasoning about our internal AWS systems.
[01:05:50] So, for example, our virtualization stack, our cryptography stack, our identity, our networking.
[01:05:59] And then, externally, customers have direct access to tools from this discipline.
[01:06:04] So, IAM Access Analyzer, VPC Reachability Analyzer, and S3 Public Access are all based on this technology.
[01:06:12] And then, outside of AWS, we see people using the space in mission-critical systems, so mission-critical situations.
[01:06:22] So, like aerospace, and railway switching, and industrial control systems.
[01:06:27] Basically, any place where failure is unacceptable.
[01:06:31] And now, we're bringing that same technology to the world of agentic AI.
[01:06:35] So, this fusion, as Swami mentioned, this fusion of formal reasoning and large language models is typically called neuro-symbolic AI.
[01:06:44] Where the neuro represents the statistical methods we're using, and the symbolic represents the symbolic formulae or symbolic logic we're using under the hood.
[01:06:58] So, how do we use automated reasoning with elements?
[01:07:02] So, one method is to use automated reasoning to verify the output.
[01:07:05] So, the output might be programs, it might be instructions for agents, or it might be Q&A pairs.
[01:07:14] If the automated reasoning tool signs off on the answer from the large language model, we're good to go.
[01:07:19] Right? But if we find a problem, we can then go push back on the language model, and it can try again.
[01:07:25] So, this creates a feedback cycle that gives us proof.
[01:07:34] Another method is to flip the equation around.
[01:07:36] So, to train the language model over the output of automated reasoning.
[01:07:43] So, for example, many model providers today use the lean theorem prover to create unbounded amounts of well-reasoned and sound training data.
[01:07:52] So, deep seek is an example of a model provider that trains using the lean theorem prover.
[01:07:59] And then the third method is to embed an automated reasoning verifier deep inside an LLM inference infrastructure.
[01:08:07] So, an example of this is called constrained coding.
[01:08:10] So, imagine an LLM that's responding to a question like, "What's the capital of France?"
[01:08:15] And it's answering with a token like B.
[01:08:19] Right? So, if we sidecar the formal reasoning representation next to the language model, we don't need to allow the language model to completely answer.
[01:08:29] Instead, it's asking the representation during the inference state to check the answers with regards to the formal reasoning model.
[01:08:39] And then we can nudge the language model instead of answering with the letter B to answer with the letter P for Paris.
[01:08:47] So, in some sense, we're shifting the reasoning left into the inference infrastructure.
[01:08:54] So, this summer, we launched Keyrow, Amazon's new agentic IDE.
[01:09:00] So, with Keyrow, developers can use natural language features to define entire applications or features in applications.
[01:09:07] So, what makes Keyrow different is its specification-driven approach to software development.
[01:09:13] So, specifications hold a very special place in my heart.
[01:09:16] They're really the foundation in automated reasoning.
[01:09:20] Specifications are how we define the behavioral properties that we want to prove or disprove of our systems.
[01:09:27] So, inspired by the success of automated reasoning and specification in AWS, the Keyrow team has brought the idea of specification into Keyrow itself.
[01:09:38] So, imagine you're building a calculator app and that needs to handle many different scenarios as well as potential error states.
[01:09:46] So, during the design phase, Keyrow might analyze the application and identify a set of acceptance criteria.
[01:09:53] So, for example, it might state when an error occurs, the app shall eventually display a clear error message in the display area.
[01:10:01] Keyrow can then convert this criteria into a specification.
[01:10:06] And then there's a lot of things we can do with that, right?
[01:10:08] So, we can use the specification to guide the language model as it attempts to find the code to implement the application.
[01:10:15] So, today, in Keyrow, you can also use that specification to generate tests that uses the structure of the specification to drive the application into interesting spaces.
[01:10:28] To help us raise assurance that the program implements the specification correctly, but also to help us pressure test the environment that it's going to operate in, our assumptions about that environment.
[01:10:39] And then it's not escaped our notice that we can also use automated reasoning to prove the correctness of the program against the specification.
[01:10:50] So, another challenge in the agentic development space is to keep up with the APIs that the code that we're generating is programming against.
[01:11:01] So, take AWS, for example.
[01:11:02] AWS is evolving faster than the model providers can update their models.
[01:11:08] And so, one way we can solve this is to develop and maintain a formal model of the APIs, and then reason about the programs that are being synthesized by the agentic tool and how they respect the APIs, so we can keep up to date.
[01:11:22] So, finally, yesterday, Matt mentioned policy and agent core.
[01:11:29] So, with policy, you get real-time, deterministic, and auditable controls over how your agents interact with outside tools and data.
[01:11:38] You can describe in natural language the actions you want to allow, and then we translate them into a formal representation called Cedar.
[01:11:45] So, Cedar is something we actually launched two years ago.
[01:11:49] It's an open-source authorization language.
[01:11:51] Its semantics are formalized in the Lean Theorem Prover, and then we've recently added an automated reasoning-based analysis to allow you to reason about the semantics of your policies.
[01:12:02] So, let's look at an example.
[01:12:05] Imagine we're building an agent to help us troubleshoot issues in production.
[01:12:10] So, what we'd like is for the agent to access production data and run diagnostics to help us figure out how to resolve issues.
[01:12:16] But what we don't want is for the agent to go make changes in production until we've agreed on what those changes are.
[01:12:23] And so, we can use natural language to write the following policy.
[01:12:26] Block any agent from performing update operations on resources in the AWS production account.
[01:12:33] So, under the hood, this type of constraint can then be automatically converted into a Cedar policy that we can run through, for example, automated reasoning to verify that it aligns with our various requirements.
[01:12:45] So, those might come from sovereignty, they might come from privacy, availability, durability, security, and so on.
[01:12:52] So, this type of combination of sound formal reasoning capabilities, together with agentic AI, represents a game-changer for making trustworthy agents.
[01:13:04] We're super excited about this work, we're super excited about the projects that are underway, and they're going to help you build agentic systems that are not just powerful, but also trustworthy and secure at any scale.
[01:13:15] Thank you very much.
[01:13:16] Thank you very much.
[01:13:26] Thanks, Byron.
[01:13:27] I hope my daughter does not learn how to use these AI agents to order a pony.
[01:13:33] Anyway, while automated reasoning helps us build agents that are accurate, we also need them to be reliable for everyday use.
[01:13:42] Reliability, if you look at it, is the level of certainty with which a task gets completed across multiple agentic runs.
[01:13:50] And when it comes to agents, businesses need high reliability when it is automating their workflows.
[01:13:58] Today, we see agents that get a task right once, but they struggle when you ask them to do it again, let alone over and over.
[01:14:06] When we think about automation, back to the early days of 2000s, the challenge of automation was handled by Robotic Process Automation, or RPA.
[01:14:17] They were designed to fill the gap between legacy systems and modern business needs by mimicking human workers.
[01:14:24] These products could log into applications and handle data entry, but they were limited in their flexibility and they would break when they encountered UI variations or complex workflows.
[01:14:36] And then came LLMs.
[01:14:38] They could handle UI variations and adapt to complexity far better than RPA did.
[01:14:44] They could navigate browsers and they could reason through problems to successfully automate workflows, even across diverse interfaces.
[01:14:53] But the problem is orchestrating these models are so complex.
[01:14:57] You need to build an error handling and backtracing because unlike a traditional script where you know exactly what failed,
[01:15:04] an LLM might keep going on down a failed path for several clicks before it recognized an error.
[01:15:11] So for many businesses, using these large language models for large scale computer use automation is simply too consuming and error prone to be practical.
[01:15:23] What a business really needed to begin with was that automation that was simple and reliable.
[01:15:29] So we asked ourselves, how can we make this easier?
[01:15:34] The answer was not just to train a better model.
[01:15:37] We needed a better model, of course, but that needs to be part of an end-to-end automation service.
[01:15:44] Tightly integrating Amazon Nova, Bedrock, H&Co.
[01:15:48] So, that's exactly what we built.
[01:15:52] Today, I'm excited to announce the general availability of Amazon Nova Act.
[01:16:03] Nova Act is a new service to build and manage fleets of agents for automating production UI workflows with high reliability, ease of implementation, and fastest time to value at any scale.
[01:16:16] It achieves 90% reliability in the enterprise workflow settings that it was trained to perform.
[01:16:23] So, what makes Nova Act different?
[01:16:27] It is because it is made up of tightly integrated components that work together to deliver this.
[01:16:34] The model is optimized to deliver best outcomes for workflow automation.
[01:16:39] Here, we started with Nova 2.0 Lite model and then post-trained it to target common enterprise workflows.
[01:16:46] So, here, the orchestrator sends screenshots of your browser to the model and then it routes commands such as clicking and typing to the actuator.
[01:16:56] It's what enables Act to use a browser the same way a person would.
[01:17:01] And finally, the SDK is your entry point to Nova Act.
[01:17:05] It enables you to build browser automation agents using simple natural language commands and they make it easy to break down highly complex workflows into straightforward tasks.
[01:17:17] This tight vertical integration across the orchestrator, model, actuators, tools, along with the SDK enables Act to better reliable than off-the-shelf models and frameworks.
[01:17:31] So, today, one of the problems is that many models are trained separately from the components that plan and execute these actions.
[01:17:39] And they can negatively impact reliability.
[01:17:43] For example, imagine you are building an agent to control a robot.
[01:17:47] Training a model separately from the orchestrator and actuators is like growing a robot's brain in a jar and then putting it into the robot after the fact.
[01:17:57] Here, we are training the brain together with the arms and the legs so that it comes out of the box knowing how to walk.
[01:18:06] To make this work, it required us to rethink how you train agents.
[01:18:11] The entire stack that the agent is trained on is consistent for both training and production, enabling us to perform end-to-end post-training that accounts for everything agents will interact with.
[01:18:25] But this alone is not enough.
[01:18:27] How you train really matters.
[01:18:30] Historically, agents were trained with something what we call as imitation learning, where the agent learns by observing and even mimicking expert behavior.
[01:18:41] This is where we started to.
[01:18:43] But the downside of imitation learning is that the agent never learns to understand the cause and the effect of its actions.
[01:18:51] So that's why we train NOVA, and for that we turn to reinforcement learning, one of the customization techniques I talked about earlier.
[01:19:00] And we focus training on workflows that enterprises encounter frequently, such as completing hardware requests, automating HR workflows, and updating customer records in a CRM.
[01:19:12] But the basis of reinforcement learning is allowing the agents to try workflows in an environment that enables them to learn from the outcomes of their actions.
[01:19:24] So to do that, we created reinforcement learning gems, where these models could train and improve, just like when you go to the gym to work out.
[01:19:35] The RL gems we created, they replicate real enterprise environments where agents can take actions and learn from their outcomes without interfering with their production environments.
[01:19:48] In an enterprise, anything you use that has a browser or UI interface could be a gym.
[01:19:54] For instance, your CRM, your HR system, or task trackers, or issue management system.
[01:20:00] In the gym, the agent is given a task or a workflow to solve, and through trial and error, learns to solve it.
[01:20:07] So we built hundreds of gyms and hand agents, run thousands of workflows on each of them in parallel.
[01:20:15] For each successful task completion, the NOVA Act got a reward, and for each failure, it got a penalty.
[01:20:23] So over hundreds of thousands of these interactions, these gyms helped NOVA Act learn patterns so it could reliably solve real-world enterprise use cases.
[01:20:34] And all of this has enabled us to create a better service and a more intelligent model.
[01:20:41] In key benchmarks like RealBench and ScreenSpot, we see NOVA Act is performing really well, as well or better than some of the best models in this space.
[01:20:52] And we are already seeing customers transform their businesses and workflows with the high reliability of NOVA Act.
[01:21:00] And if you are interested in seeing how it performs in action, check out the NOVA Act Playground or install the NOVA Act IDE extension in Kero, VS Code, or Cursor.
[01:21:11] Now, the future of agentic AI isn't agents that can do anything.
[01:21:17] It's agents we can rely on to do everything.
[01:21:21] I truly believe AI agents mark one of the most transformative steps of our time, and AWS is the best place to build and run these agents.
[01:21:32] We offer you the best tools and SDKs to quickly and effortlessly go from idea to production with agent core and strands.
[01:21:40] We provide the easiest part to customizing and fine-tuning models with Bedrock, SageMaker, and NOVA Forge, so you can maximize efficiency of these agents.
[01:21:51] We deeply invest in innovative techniques like automated reasoning and reinforcement learning so that you can meet the high accuracy and reliability needed for these modern workflows.
[01:22:03] And we didn't just stop there.
[01:22:06] When all of these things come together, the way we work is changing.
[01:22:11] We are moving from automation of individual tasks to collaboration that accelerates entire industries.
[01:22:19] Let's take a look at some examples.
[01:22:21] Yesterday, Matt announced three powerful new frontier agents.
[01:22:26] These agents are autonomous, massively scalable, and they are persistent.
[01:22:31] Able to work for hours or days in pursuit of their goals without requiring intervention or direction.
[01:22:38] Each one purpose-built to work alongside teams of people.
[01:22:43] And first is the Kiro Autonomous Agent that works alongside developers to resolve backlogs, triage bugs, and improve code coverage.
[01:22:52] Then we have AWS Security Agent that helps you build apps that are secure right from the get-go.
[01:22:59] It brings deep security expertise to review the design docs and also scan code for vulnerabilities, accelerate pen testing,
[01:23:08] and provide tailored guidance based on your company's unique security policies.
[01:23:13] And then we have AWS DevOps Agent, which integrates with observability tools like CloudWatch, Dynatrace, Datadog, New Relics, Splunk, and many more.
[01:23:23] And it integrates with your runbooks, code repos, and CI/CD pipelines so that it can engage and be your on-call assistant to triage issues, but also prevent them by being proactive.
[01:23:35] So, these frontier agents work with higher levels of autonomy and persistence, and in doing so, are able to work alongside us.
[01:23:45] Another example of these agenting teammates is Amazon QuickSuite, where AI agents work alongside business users to do research on your behalf, and to find business insights, and automate routine tasks with QuickFlows,
[01:24:02] and create multiagent workflows, and create multiagent workflows, all without ever needing to write a single line of code.
[01:24:10] Our vision for the future of work is that every person in an organization have intelligent, agenting teammates that amplifies their capability and eliminate all the tedious tasks so that they can focus on the work they love the most, and unlock insights that drive real business transformation.
[01:24:31] To share more of what we are doing in this area, I'm excited to welcome Colleen Aubrey, our SVP of Applied AI Solutions.
[01:24:40] Thank you, Swami.
[01:24:50] I believe that over the next few years, agenting teammates will be essential to every team, as essential as the people sitting right next to you.
[01:25:02] They will fundamentally transform how companies build and deliver for their customers.
[01:25:13] At the heart of this transformation is AI woven into daily operations, embedded in every workflow, and as natural as checking your phone in the morning.
[01:25:24] But when do we pull the trigger?
[01:25:26] Do you wait until every question about AI is answered?
[01:25:31] Every risk mapped out?
[01:25:33] Every unknown eliminated?
[01:25:35] That's the trap, isn't it?
[01:25:36] We convince ourselves that clarity equals confidence, that predictability equals efficiency.
[01:25:45] But luckily, here's what I've learned of building at Amazon.
[01:25:49] Transformation and agility, they're not opposites, they're actually partners.
[01:25:54] And real efficiency isn't static, it's alive and it adapts.
[01:25:59] Now, making meaningful change is also not about automation.
[01:26:04] Automation assumes that the current approach is optimal.
[01:26:08] That the ultimate prize is less effort.
[01:26:11] But many times, the status quo is not optimal.
[01:26:16] The real prize of AI is new products, new services, better customer experiences, and new business models.
[01:26:25] Not less effort.
[01:26:26] But there's a journey to get there, and we all know this is a journey that is accelerating.
[01:26:32] It's possible that the biggest limitation to transformative change is not a technical limitation.
[01:26:38] It's our own ability to reimagine how we work, how we collaborate, and how we make AI part of the team.
[01:26:46] So, what does it look like to have an agentic teammate?
[01:26:51] One of the first places we've put agents to work as teammates is in Amazon Connect.
[01:26:57] Connect is an easy-to-use, cloud-based, omni-channel, AI-native customer service application.
[01:27:03] It's based on the same technology we use within Amazon to power our own customer service.
[01:27:08] And it's used around the world to power billions of customer conversations.
[01:27:14] Customers like DXC Technology.
[01:27:16] They're a global provider of IT and consulting services.
[01:27:19] They help companies to modernize their systems.
[01:27:22] They use Connect to intelligently determine which interactions to route to which people.
[01:27:27] And then empower those people with real-time assistance.
[01:27:30] It pulls in from internal knowledge bases while respecting the access and security controls.
[01:27:36] This enables DXC to deliver very personalized context-aware support across every channel.
[01:27:43] What we see across many customers is that the center of gravity for customer experience is the same.
[01:27:50] It's starting in the contact center, but let's be clear.
[01:27:53] The contact center have expanded beyond just the interactions.
[01:27:59] Today, they're about the operational backbone for customer relationships.
[01:28:03] And the agentic capabilities of Amazon Connect are there to help through each challenge.
[01:28:08] So, let's see it in action.
[01:28:11] Many transactions involve a bank or a credit card or a digital payment.
[01:28:17] And as a cardholder, one of the sort of most concerning moments is when you see a transaction on your account that you just do not recognize.
[01:28:24] It can happen to anyone, even me.
[01:28:28] Earlier today, I saw several transactions on my account that I didn't recognize.
[01:28:33] So, let's see if we can call customer service and see if they can help.
[01:28:43] Good morning, Colleen.
[01:28:45] Before we get started, let me verify your identity.
[01:28:48] Can you tell me your passphrase?
[01:28:50] Sure, it's "show me the money."
[01:28:53] And can you tell me one of your recent transactions?
[01:28:57] Yeah, let me think.
[01:28:58] This morning, I got a Cortado at Ili.
[01:29:01] Thanks, Colleen.
[01:29:03] I was able to verify your account.
[01:29:05] I see in your account notes, you're traveling right now.
[01:29:08] Are you having problems with your card?
[01:29:10] Yeah, so this morning, I saw several transactions that I just don't recognize them.
[01:29:16] I'm sorry to hear that, Colleen.
[01:29:18] Which transactions?
[01:29:20] There was a bunch at gas stations that I've never been to.
[01:29:24] But also, I'm not even driving while I'm here.
[01:29:26] Okay, I see them.
[01:29:29] I am locking your card so we can start an investigation.
[01:29:33] Whoa, wait a second.
[01:29:34] I'm at a conference.
[01:29:35] This is the only card I have with me.
[01:29:36] I need to use it while I'm here.
[01:29:39] Not to worry.
[01:29:40] I see you've added this card to Apple Pay.
[01:29:42] We will only block transactions with your physical card.
[01:29:45] So, you can continue using Tap2Pay on your phone and registered devices.
[01:29:49] And I'll also expedite having a new card sent to your home address.
[01:29:53] Okay, great.
[01:29:54] That works.
[01:29:55] But what about these transactions?
[01:29:56] How do they happen?
[01:29:57] Is there some fraud going on?
[01:30:00] I'm taking a look now.
[01:30:02] I'll need to review the details to see if I can identify a pattern.
[01:30:06] I'd like to connect you to my colleague while I do this.
[01:30:09] I'll share what I find while they talk with you.
[01:30:11] Let's see if we can get to the bottom of this quickly.
[01:30:14] Okay, great.
[01:30:15] Let's do that.
[01:30:18] Hi, Colleen.
[01:30:19] I'm Hector.
[01:30:20] I'm the investigator assigned to your case.
[01:30:23] The agent gave me all your information.
[01:30:25] Oh, thank God.
[01:30:26] I thought I was going to have to repeat all of it.
[01:30:28] Nope, don't worry.
[01:30:29] And let me just say I'm so sorry.
[01:30:31] This is happening to you, especially when you're at a big conference.
[01:30:35] Thanks, Hector.
[01:30:36] So, what happened?
[01:30:37] Are my other accounts at risk?
[01:30:41] Let me pull out the transactions and take a look.
[01:30:43] Okay.
[01:30:44] That sounds great.
[01:30:49] All right.
[01:30:50] So, analyzing and verifying fraudulent activity is a process that often takes hours or days.
[01:30:57] Using my agentic teammate to help me, I'm going to do all of this in just a few minutes.
[01:31:02] So, let's get started.
[01:31:09] All right.
[01:31:10] Let's see this.
[01:31:11] All right.
[01:31:12] I see the agent has already flagged the suspicious activity and analyzed them all.
[01:31:16] And it found that the geographic pattern of the transaction seems suspicious.
[01:31:22] So, let's take a look.
[01:31:25] Transactions on a map.
[01:31:26] And I can see these transactions happen all across the U.S.
[01:31:31] And as far as I know, Colleen can't teleport.
[01:31:35] So, I agree this looks super sus, as my daughter says.
[01:31:40] The agent also tells me this pattern is consistent with a scheme card.
[01:31:45] Let's see if we can verify this by asking the agent to look for patterns across other cases.
[01:31:52] And in just a few seconds, the agent was able to confirm likely fraud.
[01:31:57] As I already suggested, our standard operating procedure for this type of activity, which is to create a police report and share the details.
[01:32:05] So, I'll go ahead and do that.
[01:32:08] Great.
[01:32:09] Okay.
[01:32:10] So, we handled the fraud.
[01:32:11] But that's one problem solved.
[01:32:13] And Colleen was also worried about the other problem.
[01:32:16] So, let's check those out.
[01:32:19] Next, I'll switch to Colleen's to connect agents' building experience.
[01:32:23] So, I can create custom agent that will be able to watch Colleen's account for any suspicious activity.
[01:32:29] I just give the agent a name, type, configuration template, and a quick description.
[01:32:37] I can defend agent's behavior by providing a simple prompt.
[01:32:41] Next, I need to apply guardrails for the agent.
[01:32:44] And that's it.
[01:32:45] Let's get back to Colleen and tell her what we did.
[01:32:48] Alright, Colleen.
[01:32:49] It looks like your card was most likely skimmed at an ATM you used at the airport.
[01:32:54] What?
[01:32:55] Really?
[01:32:56] Yeah, really.
[01:32:57] But analyze all your other accounts, and good news, there's nothing abnormal.
[01:33:01] For extra peace of mind, I set up automatic health checks on all your accounts that will alert you via a secure message if anything suspicious does happen.
[01:33:10] Okay, great.
[01:33:11] Thanks.
[01:33:12] I travel a lot for work.
[01:33:13] This is super frustrating.
[01:33:14] Is there anything we can do to stop this going forward?
[01:33:17] Yes, super frustrating, but let me see what else I can do.
[01:33:22] I have AI analyzing your full account and transaction history to see if there's anything else I can offer you.
[01:33:29] And it looks like you qualify for a secure travel account that comes with a more secure card and won't cost you anything.
[01:33:36] Okay, sounds promising, but what about travel benefits?
[01:33:40] Well, this new account comes with great travel rewards.
[01:33:44] From transactions, I found that you have a reservation at the MGM Grand later tonight.
[01:33:49] Ah, yes.
[01:33:50] La Tele.
[01:33:51] I love that place.
[01:33:52] Have you been?
[01:33:53] Me?
[01:33:54] No.
[01:33:55] I don't go to restaurants.
[01:33:56] I can't pronounce the name, but you know, sounds nice.
[01:33:58] Okay, so let me get this straight.
[01:34:00] More security, more perks, and it doesn't cost me anything extra.
[01:34:05] Yep.
[01:34:06] That's right.
[01:34:07] Okay, that sounds great.
[01:34:08] Let's do that.
[01:34:09] All right.
[01:34:10] Well, your new card information is available in your app so you can configure it with Apple Pay.
[01:34:17] Enjoy your day and your lovely dinner, Colleen.
[01:34:19] Okay, so I don't suppose anyone had join a customer service call on your bingo card today.
[01:34:28] That's okay.
[01:34:29] Being able to make a meaningful difference in customers' lives and build relationships that, you know,
[01:34:35] makes that human-agent collaboration so powerful.
[01:34:38] And as we work across multiple modalities with agents, our need for seamless escalations and human interaction still exists.
[01:34:48] And it's exactly why we keep investing in capabilities that make people and AI agents true partners.
[01:34:55] In fact, earlier this week, we launched eight new agentic capabilities in Amazon Connect.
[01:35:02] From NovaSonic integration that enables customers to problem solve through spoken conversations that sound natural,
[01:35:08] like the one you just heard, to agents that make real-time recommendations based on the customer conversation itself
[01:35:14] and automatically suggest next steps, then to AI-powered predictive insights that combine clickstream data, customer profiles,
[01:35:23] to make highly personalized suggestions.
[01:35:27] Over the next few years, human AI teams will fundamentally rewire how work gets done.
[01:35:34] They'll change how we build, how we ship, and how we serve customers.
[01:35:40] And here's what excites me the most.
[01:35:42] This isn't just about doing the same things faster.
[01:35:45] It's about unlocking capabilities we couldn't even imagine.
[01:35:50] Thank you.
[01:35:51] Thanks, Colin.
[01:35:51] Connect is a great example of how agents and humans can work together to solve problems efficiently.
[01:36:09] So everything we are shown today and everything we are building, it's all just the beginning of what's possible with agentic AI.
[01:36:17] And when I think about the innovation that will happen between now and when we are back here next year,
[01:36:23] I can't help but be optimistic.
[01:36:26] In every step, we are making it easier for anyone to build and use these agents.
[01:36:31] But the technology we build is just one part of the equation.
[01:36:36] The other more critical part is you.
[01:36:39] It's the community of builders inventing and building these incredible things.
[01:36:44] In October, AWS Builders Partners and Schools in West Java, Indonesia, had a goal to create 2,000 Gen AI apps.
[01:36:53] They far surpassed this, setting a Guinness World Record for most apps made in an on-site event
[01:36:59] over 15,000 Gen AI apps.
[01:37:03] Almost all of these apps were dreamed up by high school students
[01:37:07] who were able to bring them to life with the help of their teachers
[01:37:11] and a few non-profit staff.
[01:37:14] And on September 8th, we launched the AWS AI Agent Global Hackathon.
[01:37:19] The challenge was simple.
[01:37:21] Build autonomous AI agents that solve real-world problems.
[01:37:25] This year, we saw 9,500 developers from across 127 countries participate in this global hackathon.
[01:37:33] And they were challenged to build, develop, and deploy working AI agents using Bedrocks,
[01:37:39] SageMaker, Trans, Agent Core, Kero, and many more.
[01:37:43] And we received 625 agents covering use cases from fraud detection, to infrastructure monitoring, to agriculture health monitoring.
[01:37:53] Showcasing how AI agents are being deployed across virtually every industry out there.
[01:37:59] But the winning agent was built by Ahito Nelson.
[01:38:03] Ahito is from the country of Timor-Leste, where waste management is an incredible challenge.
[01:38:09] The capital city of Dili produces over 300 tons of daily waste, yet more than 100 tons go uncollected each day,
[01:38:19] clogging the drainage systems, causing devastating floods, and creating serious health assets.
[01:38:25] Ahito built Echol Lafayette, an autonomous AI agent that transforms every smartphone into environmental monitoring tool,
[01:38:34] empowering citizens to become guardians of their community's cleanliness.
[01:38:40] He used Amazon Bedrock with Novapro for multimodal analysis.
[01:38:44] Titan for semantic similarity search.
[01:38:47] Agent Core with code interpreter and web browser for automation.
[01:38:51] And AWS LightSail with S3 storage.
[01:38:54] Then citizens can now photograph waste issues,
[01:38:57] and the AI instantly classifies waste types, assesses severity,
[01:39:02] and then identifies hotspots.
[01:39:04] So they're giving authorities the tailored guidance they need to prioritize cleanup efforts.
[01:39:10] This year, we also had a fantastic initiative called Road to Reinvent,
[01:39:15] where teams scored on buses with a challenge to build an authentic AI app en route to Vegas,
[01:39:22] and complete for the grand price of $50,000.
[01:39:26] And finally, I also want to congratulate the winners of this year's AI league,
[01:39:32] an AWS program where builders compete by customizing and fine-tuning LLMs and agents for specific tasks.
[01:39:41] And in 2026, we are going even bigger, increasing the grand price to $50,000.
[01:39:47] You can check them out in export today.
[01:39:51] The energy and the excitement we are seeing around authentic AI is truly incredible.
[01:39:57] The winner from this year's championship and Road to Reinvent are here with us today.
[01:40:02] Please stand and take a bow.
[01:40:16] Now, I want you to think about the feeling again.
[01:40:20] That moment when you first made something work.
[01:40:23] When you first felt that rush of creation.
[01:40:26] That feeling is not just a memory anymore.
[01:40:30] With authentic AI, we are living in that moment every single day.
[01:40:35] We are at a point where anyone with an idea, whether you are cleaning our oceans,
[01:40:41] unlocking the mysteries of the human brain,
[01:40:44] or solving challenges that we haven't even discovered yet,
[01:40:47] has the freedom to build it.
[01:40:49] The freedom to move from concept to impact at an unprecedented speed.
[01:40:54] To tackle problems that once seemed impossible.
[01:40:58] The freedom to create without limits.
[01:40:59] So, I want you to think about how you will use AI agents as you take on your next challenge.
[01:41:09] Let's build this amazing future together.
[01:41:11] Thank you.
[01:41:12] Thank you.