What I love about technology is that it can help us do things that we could have never imagined. For instance, I grew up in a rural part of India. I didn't grow up in the city. I didn't come from an affluent family. In fact, we didn't have a computer when I was growing up. My middle school and high school had one computer that the entire school shared. I got access to 10 minutes a week, maybe 20 minutes max, for me to actually use a computer. That means I got to make every second count, and every second was precious because I wanted to learn how to program. With only 10 minutes to go, it wasn't an obvious choice or an easy one. I didn't have all day to try out my code. In fact, I had to be a human compiler to detect these syntax errors ahead of time. I fell in love with this problem solving that came with this and went on to actually study in the top college in my state, College of Engineering, Guindy, and was the first generation in my family to go to college. Eventually, I went out to get a PhD in Vrije Universiteit in Amsterdam. One funny anecdote, at my university, you had to have two people standing by your side while you are defending your thesis. In case the defense kept going on and on, someone needs to stand in if you need a break. Here in this picture, I asked my brother to be one of them. He knew almost nothing about my PhD dissertation and was terrified that I would step away as a joke, but I didn't.

Eventually, I got a job at Amazon. You've got to remember, this was 20 years ago. And during that time, and I was, actually I distinctly remember, calling my mom and telling her, saying, like, "Mom, I got a job in Amazon." And I still remember my mom's reaction when I told her. She was certain I was going to waste my time, PhD, by joining an internet book company, because that's what Amazon was at that time. But at Amazon, I got an opportunity to build amazing things. And what became eventually AWS. And I got to build technologies like DynamoDB, SageMaker and Bedrock, which are the underpinnings of many of the modern applications we use today. And now, if I look back, it all started with the 10 minutes of access I had to that computer that wasn't even mine. It opened up worlds to me that I could have never thought that was possible. And now, as the VP of agentic AI at AWS, when I think about how agents are going to transform everything, I can't help but be optimistic. Today, I’m going to talk to you about AI agents, what I think will be one of the most transformative technology shifts of our time. And we will talk about what they are and what are the milestones they need to achieve before we can trust them and make it an integral part of our daily lives. And also talk about how they will change everything.

So first, what are AI agents? AI agents are these autonomous software systems that leverage AI to reason. They plan and they adapt in pursuit of user-defined goals. They complete tasks on your behalf of humans or other systems. These AI agents can sense and interact with their digital environment, converting these high-level objectives into executable steps. And constantly they learn and improve their efficiency over time. Today, agents are being used for everything, right from software development to drug discovery to precision agriculture to many more. Their ability to use and manipulate interfaces in their digital environment, the same way we as humans do, dramatically lowers the bar for use cases like building applications. You no longer need rigid application specifications and then break it down into complex software projects. Now you have the possibility to just state your goal and let the AI agents figure it out.

But not everything is an agent. For example, imagine you are a researcher in a lab. You're sitting down at your computer and tell the AI that you want to run some experiments to explore a new protein. It responds telling you something like, "Great, let me actually propose the six experiments you can run." Now that's not an agent, that's a chatbot. But with agents, what you get is when you give them a goal, they can plan. They can write code, they can use the tools to build the experiment for you. They will synthesize your results, and they will reflect on failures. And they will look for ways to constantly improve their efficiency over time. The work that you might take for a week or more to research and build the plans for these experiments can now be done in hours or even minutes. Your role now becomes more of a trusted advisor, where you are steering these AI agents towards actually execution and, in many ways, like peer reviewing a colleague’s book. With AI agents, the barriers to creating something will now lower. Challenges, like, I don't have a particular skill, or I don't have enough resources or headcount to do this project are going to start to go away. The future we will share will be shaped by those with the ability to think big and even dream bigger.

But we are not there yet. In fact, there are three milestones these AI agents need to achieve before they fundamentally change how we work and how we live. The first is how we build software. So much of our world is digital. In fact, in this room alone, on all the devices you have in this room, there are probably hundreds of applications, if not more. In our daily lives, on a constant basis, we carry the works of tens of thousands of software developers, if not more than, like, hundreds or thousands of software developers. So now, when you think about it, for AI agents, before they can even reach the masses, they need to reach builders. And that means, if they are going to survive, those builders need to find the agents to be useful and interesting. This goes beyond the tools that these developers use on a daily basis. They are already becoming agentic, but what needs to change is how easy are these agents to build. The bigger shift is in changing how we conceptualize effective agent architectures.

Now take a look at this problem of how software developers or builders build applications. Today, as developers, they have a bunch of choices they have to make as they are building these applications. Many of these are implementation details, like which server or which compute option do I need to choose for hosting this website or building this mobile app? If you have never had to make this choice, there are a lot of options to decide, like, how to host your website or application in the cloud. For example, in AWS for a builder, if they want to host a mobile app or website, in one of our services called EC2, we offer something like 850 compute options for them to choose from, and that is not even one compute option. There are even more. And now, as we move towards the agentic era, developers will be able to shift their focus into what they are building instead of worrying about how they are building. That means decisions like which compute to choose become less relevant. In fact, AI agents are going to automatically enable us to pick those things for you.

Now the next milestones agents are going to reach, and they must reach, is trust. Without trust, none of these capabilities of agents are going to really matter. But how can you trust an agent? The reality is that we are still in very early days of agentic AI. We know agents are imperfect, and they will make mistakes. Yet even in simple tasks, we have an uncompromising need for perfection. The good news is that agents are not reaching into some magical ether to make things happen. The systems, tools and the environments that these AI agents are using have well-understood specifications on how they work and what they should be doing, so they can actually be mathematically proved if a system or program obeys its application specifications the way they are intended to. And this technique is called automated reasoning.

Automated reasoning is a field of computer science that attempts to provide assurance if a system is behaving exactly as it is expected, based on sound mathematical logic. Its roots go all the way back to ancient Greece, where Aristotle was the first logician to attempt a systematic analysis of logical syntax. Today, automated reasoning is the algorithmic search for proofs in mathematical logic, and can be used to make sure that the agentic reasoning is accurate. To do this, you need to know precisely what each agent can do.

So now let's take a look at a very small example. At AWS, one of the first agents we built was called Amazon Q. Among other things, Q was built to help software developers build software applications. We were really excited. We were already imagining all the amazing possibilities Q can do, even in its prototype, because it was going to be as smart and capable as our best software developers. We thought it's going to accelerate our road map and obliterate all our backlogs. But there was a problem. The first prototype we built was more like me when I was an intern in Amazon. They were eager and error-prone. They were hallucinating API calls. We had to fix it.

So how did we go about it? We formalized all the API specs into mathematical model so that every time Q generates an API request, an automated reasoning solver first verifies saying like, "Is this a valid request?" If the solver finds an error, it communicates back to the agent saying like, "Hey, I think you got it wrong this way. Can you now restructure your code?" So now it gets fixed even before requiring human intervention. This back-and-forth communication creates what I call as a neurosymbolic feedback loop that is completely transparent and enables us to mathematically prove that the action an agent can take is going to be correct even before it is taken. And it does it faster than you can blink: 100 microseconds or less for 95 percent of use cases. Now, this is just a small start, but we believe combining agentic AI and automated reasoning will help agents become trustworthy to reach widespread adoption.

Now, if we stopped here, we would have an incredible developer experience where every software developer in the world can build amazing, trustworthy agents. But agents can't change everything if it only targets a small subset of population. Across businesses, there are wide variety of people and most have never written a single line of code. The final milestone is to enable anyone to build agents. Here is an example. Imagine if you only had two minutes to recap everything you heard in the TED conference today and now you had to summarize it. For many of you thinking, "You know what, I'm going to just talk really, really fast and I can do it." That's not going to cut it if I tell you you have to use the clips that you saw today to create your two-minute summary. How long do you think it will take you to create this perfect two-minute story?

Now that is the exact problem we faced in our Amazon Prime Video, where an effective recap of a Prime Video series can take weeks to produce and is very expensive because everything, from creating the story arc to selecting scenes, is manual. Cinematography experts are not usually the master coders, but we introduce agents to help streamline the process, breaking the workflow into three phases: observation, reasoning and action. Now, in our first phase, what we call as observation, we ask AI agents to understand what's happening in the video. They need to produce a rich and detailed observation and understanding about every aspect of the shot, scene and the entire story, so that we can define a story arc and select the right scenes.

Then we move to the second phase, what we call as reasoning. Here, what we can imagine is the agents are saying, like, "With what I know, what do I need to do?" Reasoning layers on top of observation. So for example, we want to generate a voice-over narrative for recaps. We can ask the reasoning agent to generate the script by collaborating with the observation agent. Then, the final step is what we call as action. In this phase, now what you are bringing in are the trusted experts who are going to work with these AI agents to help finally recap the story. Now, if you go back to your two-minute TED recap, how much easier would it be for you with this task if you had these powerful AI agents? The power of human and agent collaboration is that it frees us up from being bogged on by the drudge work, and enables us to do these amazing things and creating things based on exactly what we love.

But in Prime Video, they were using agents. So how do we get to a place where anyone can build agents? In fact, the frameworks to build agents are already getting simplified day by day. Any application developer who knows how to write a code in Python can now build a pretty useful agent already. And now we are also starting to see, like, not just from AWS, but everywhere around, we are building these agentic cloud infrastructure that makes it super easy to go from proof of concepts to production. But those alone are not enough. We need to expand the pool of people who can build AI agents. To get there, the interfaces to build agents must become familiar to business users as well. The way we think about building and training agents must also change. Smarter models are great, but a world-class scholar that doesn't take any action or that is ignorant to the way we do things, isn't helpful. We need agents that are ready for the real world. We will need to create worlds for agents to play with and to improve the next generation of digital twin.

And once we are done [with] all of that, what happens? If we get it right, these agents will become invisible, but they will help us do incredible things. In the next few years, we will see agents that give rise to more companies faster than ever, where success is determined by your ideas and your ability to describe what you want to build. We will see more medical breakthroughs, and you are going to see way more discoveries. And with all of this, what makes me so optimistic is that the future we will have with agents will be ultimately built by you. Your 10 minutes are coming. What will you build? (Applause)