AWS VP Swami Sivasubramanian on agentic AI going into production — Bedrock request volume in Q1 exceeded all prior years combined

Speaker 2: He is the VP of AgenTik AI, and he's here with us on the TBPN UltraDome. How are you

Speaker 1: doing? What's going on?

Speaker 3: I'm I'm great. Hey. Thanks for having me.

Speaker 2: Thanks for hopping on. Why don't we start with a little bit of an introduction since it is the first time you're on the show? Tell us a little bit about, how you fit into the AWS organization, what you're working on, and then we can go into AgenTik AI.

Speaker 3: Yeah. I mean, I'm the vice president of AgenTik AI. I've been with AWS now for, like, twenty years. Mhmm. Created various things Overnight success. From, like, our data for Hold on.

Speaker 2: What was it like twenty years ago? I feel like AWS is barely 20 years old. Were you, like, the first person on the team? You and Andy were hanging out and racking servers?

Speaker 3: Kind of. Actually, literally, AWS was would fit in a single conference room if you can't That's believe amazing. And I mean, I joined as an intern. My internship project was to build a database called Dynamo. And yeah. I mean, that turned out to be a bit of a success, so they gave me a job back.

Speaker 2: So That's great.

Speaker 3: And then I started the database joined the database team, started bunch of databases. Yeah. Then moved on to analytics and started AI for AWS.

Speaker 2: Okay. Okay.

Speaker 3: And now Yeah. Now that you see all the things with Bedrock and SageMaker, all the AI tools, now as the world is moving more agent tech, I kind of spun myself out of what I was doing Yeah. To focus all in on AI agents. So that's kind of my background.

Speaker 2: Yeah. I imagine that, you know, the the the typical database business, the analytics business, all of that is growing, but it it's it's being driven by more agentic AI use. Talk about the new releases, the new launches, and I wanna know how do we compare and contrast continuum from context? Where how how do companies use the right tool for the job here?

Speaker 3: Yeah. I mean, first of all, we live in some amazing times with AI agents because the art of possible has fundamentally changed. But when you look at it even where you see, actually, enterprises are struggling today. First of all, on putting agents to work Yeah. They are still stuck in this world of, I call it, the walled gardens, where they are stuck into bottlenecks where if I go on vacation or even when I go to sleep and wake up, I get signals from, like, my Outlook to Slack to various other messaging tools. So the time I spent to even catch up used to be, like, crazy, like, least an hour before I know I get to my job. Mhmm. So why why was it that case? It's not like we didn't have AI assistance in every one of these tools. Mhmm. But what was missing was it didn't have the context to fill in the spaces between these various tools. So that's why today, actually, we launched something called Quick Autonomous Agent. It actually works on your behalf. It gathers all the context across tools so there is no more walled gardens, but it still has something enterprises know. It has built in governance and security. That's why the likes of GoDaddy to NBA to everybody is using it. You are showing it live today, and the art of possible, even when we demoed it live, people couldn't believe it because now you can actually catch up and generate a PowerPoint deck in a matter of minute to a detailed dashboard when you talk to your data lake and generate insights, all these things are changing. That is one example of bottleneck we are removing. The second one I talked about was in the world of security. Security is a big area of our intersection of how software is getting built and security is changing in a big way because, yes, the rate of change in how you can generate code is becoming very, very cheap. Mhmm. But, also, the response security is also becoming big. So speed and security are almost at a trade off in every organization. So what we launched today is a service called Continuum. In many ways, the approach behind it is make security not a discrete thing companies do. It's like how antibodies actually fight viruses. Always on continuously fighting is the way we thought about it as well, as an example.

Speaker 2: Yeah. I wanna dig deeper into those, like, continuously running AI agents. I think when people think agentic AI, they go to coding agents because everyone's had the magical experience with those and seen the impact and the spend and the token maxing that happens when you deploy AI agents across a huge workforce of knowledge workers, heavily software engineers. But I'm I'm more interested in in the less understood agentic workflows that are happening maybe deeper and autonomously in a business. Yeah. And I'm trying to get to a world where I have that case study at the tip of my tongue for every time I book a flight, it runs an agentic workflow to make sure that my seat on the plane is correct. Or I don't know. I I I don't know how to contextualize it but I know it's happening and I know you're seeing it. So walk me through some of the agentic workflows or agentic AI work that's happening even if no one shows up to the office. It's a it's a company holiday and there's still tokens being spent. What does that look like? What is the shape of that work?

Speaker 3: Actually, I mean, that is one of the key things that we are going to start seeing explosion of agents in a big way. Even today, for instance, you, in your summit, the CIO of Southwest Airlines talked about how they are building agents on top of actually agent core, which is our core agent platform service. And they were able to build these amazing new capabilities to do things like crew planning and various others. Mhmm. These are not like very humans are just generating codes, but how that's where they are going. Three and four instance are using agents to now be able to do more efficient sales call planning. GoDaddy is able to actually leverage, some of our quick agents to automatically reduce something like fifteen thousand, hours of manual work that they are able to do. These are, like, examples of agents. Again, not created just for purely creating software, but these are agents running under the hood, and they are able to produce value. And this is gonna be where the next big agentic workload is going to happen in many ways, and that's what we are excited.

Speaker 2: What what is demand like for model routing these days? AWS is fascinating because you've sort of maybe it's brilliant by design, but sort of the Switzerland of AI in the sense that there's partnerships with Anthropic, OpenAI. There's a whole bunch of other foundation models that are available on AWS. What is demand like? Obviously, different teams have different preferences and cost tradeoffs and they might say, I want to be with this particular Claude model or this particular GPT model. Other companies might just say, I never I never want to be caught using a frontier model for something that doesn't need that doesn't need frontier So optimize it for me. Is that is that something that you're seeing demand for already?

Speaker 3: Yeah. I mean, first of all, I think for the industry itself, when I started Bedrock, one of the things we were different compared to everybody else's. We always had this belief no single model will rule the world. This was not a popular belief at that time, but but now it seems so obvious. But then amazing thing is now that you are the Switzerland as you actually speak, in q one alone, the adoption for our Bedrock, Generative AI LLM platform, what we are starting to see is in q one alone, the amount of workload that, request rate that we are seeing is greater than all of the previous years combined. That is the the amount of growth we are seeing.

Speaker 1: Crazy. And,

Speaker 3: it seems to have no signs of slowing down. It's because the rate of actually adoption, especially as people are moving from proof of concepts to production, is just accelerating because they are seeing real outcomes. And that is one of the reason why so much of innovation is going into things like Bedrock and agent core, not just on the model front, but also how we are supporting the things like disaggregated inference and all the agentic capabilities such as launching new context. So these are the capabilities that's gonna make actually enterprises move from just cute prototypes that are sitting in by the demo to their, execs to actually production to see real value. Yeah. And this is why I do think there is so much upside in the future.

Speaker 2: How do you think about AWS as a sort of like a thought leader for for enterprises rolling these features out? I'm just thinking about, like, we're going through this transition of like there's a new technology. You need to use it effectively. I feel like a big piece of why AWS was so successful was because it powered amazon.com. And if it's good enough for amazon.com, it's good enough for my business or whatever I'm doing. And it feels like now there's an opportunity for AWS to sort of set the agenda on what it's not token maxing. It's not token minimizing. It's token optimizing. It's ROI maxing or something like that. But how are you communicating to your team, to the teams within AWS about how to use AI effectively? Sometimes you just wanna get everyone up to speed, just go and test everything. And then sometimes you want to be, you know, setting a tone that a company that says, hey, look, we don't have a lot of free cash flow here to throw at this new hot technology. If we're going to use something, it has to deliver ROI almost immediately. How are you carrying that through your organization and into the companies that partner with AWS?

Speaker 3: I mean, it's a great question because if you look at the industry, and this is true even at Amazon, we first started with everyone should get familiar with these tools. So we were trying to get everyone to adopt so that they know how to use these tools. Mhmm. Now that they have become so familiar, we are now making all the tools available between, like, Kira or agentic code Mhmm. Putting solution to Cloud Code to Quake as our business user productivity solution. Within Amazon itself, tools like Quake is used by hundreds of thousands of people, let alone Keyroad to many others equally on the software development. Now we are come to a point where we internally and this is true for our external customers such as Delta and Southwest and many others as well as we give organizational breakdown. Within Amazon, we say, hey. Every VP under that team, how much usage is happening? What is their cost breakdown?

Speaker 2: Yeah.

Speaker 3: And then we roll it out. This is true for Kiro, and pretty soon it'll be true for Quik. And within Bedrock, which is our Gen AI platform service, we actually integrated with AWS Cost Explorer so that we can actually have people spare I mean, monitor their spend Yeah. In terms of our at a per user level, per model level, and a per organization level too. The moment you have visibility, then you know Mhmm. Where people are spending, then you can understand ROI. In fact, the interesting thing what we saw was that my teams and there are a few teams in Amazon. I call them frontier teams. These are teams that, don't see 50% productivity improvement. They see 10 x to 20 x productivity improvement. And when you see their bill on token, it's not actually, like, a $100,000 a month. It's actually even in the keynote credits, it's something like around 2,000 to $3,000 a month. Sure. That's not that much for 10 x to 20 x productivity. Yeah. But I do think this is gonna keep going as the world becomes more and more agent tech. Because once they move to coding to be just launched our release agent because you increase code velocity. Everyone talks about coding, but this is the myth. Suddenly then, everything gets stuck because you need to deploy it to real world. And deploying it to real world is a bigger challenge because you can't just increase the volume by 10 x a year, and then deployment is going at one x. That means everything becomes a technical debt. So we made that agent tech, and then we made pen testing agent tech. Mhmm. So I almost jokingly call it saying, like, hey. You need to ride it correct, ship it fast, and constantly keep it modern. And the only way to do it is you need to have all of it be done by agents, and that's exactly how we are now moving towards.

Speaker 2: Makes sense. Well, thank you for coming on the show. Congratulations on the launches.

Speaker 1: And Yeah. Great to

Speaker 2: meet you. Talk to you soon. Have a great rest of your week. Bye. Hey. We'll talk to you soon.

Speaker 3: Again forever. Of course.