AWS re:Invent 2025 | Swami Sivasubramanian, AWS

Clips
More from AWS re:Invent 2025

Swami Sivasubramanian

VP, AI and Data

AWS

play_circle_outline Agentic AI seen as a major technology transformation in software development.

play_circle_outline Revolutionizing Productivity: Kiro's Agentic IDE Automates Tasks and Enhances Software Development Efficiency

play_circle_outline Emphasis on the importance of agentic teammates learning team preferences for better collaboration.

play_circle_outline Building Trustworthy Agents: The Future of Enterprise Integration and Agentic Applications in Business Operations

Info
Transcript

Swami Sivasubramanian, AWS

Swami Sivasubramanian

VP, AI and Data AWS

In this interview from AWS re:Invent 2025, Swami Sivasubramanian, head of Agentic AI for AWS, joins theCUBE’s John Furrier to discuss the seismic shift toward agentic AI. Sivasubramanian frames this evolution as one of the most significant transformations in tech history, positioning it not just as a productivity booster but as a fundamental reinvention of software development. The conversation highlights the launch of Frontier agents and NovaForge, noting how AWS is moving beyond basic code generation to address the full development lifecycle. Sivasubramania... Read more

explore Keep Exploring

What is the significance of agentic AI and how is it expected to transform technology in the future? add

What are the benefits of using virtual agentic teammates in software development and how has the approach transformed project completion times? add

What are the limitations of current AI assistants compared to more experienced team members? add

What is anticipated to be the future development or trend following the upcoming re:Invent event? add

bolt Powered by CUBE AI

Swami Sivasubramanian, AWS

search

John Furrier

>> Welcome back on theCUBE's live coverage of AWS re:Invent 2025. I'm John Furrier, host of theCUBE. It's our 13th year covering re:Invent. We've seen the evolution of Amazon Web Services and cloud from abstracting away servers and infrastructure to building higher level services. But now that the ground is changing, but the game is still the same, abstracting away data and work, you're seeing agents really starting to form the next view of this next level of value. Back on theCUBE as a tradition, Swami's back, head of Agentic AI for AWS. Great to see you.

Swami Sivasubramanian

>> Hey, glad to be back.

John Furrier

>> I love to see you in the hallways here. I just grab you, pull you in. It's called the hallway grab in theCUBE vernacular. Thanks for coming on.

Swami Sivasubramanian

>> Oh no, my pleasure.

John Furrier

>> So keynote today, smashing. Again, all in on agents.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> We saw each other at the HumanX Conference as well. What has progressed in your mind? Because this year, a couple, three things jump out at me. One, the AI factories positioning basically speaks to the large scale.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> Infrastructure. Great. Check the box.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> Trainium3, better performance.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> We're seeing all the industry on the CapEx side.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> But the NovaForge and Frontier agents really jumps off the page as a game changer.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> Explain why that's so important and why now.

Swami Sivasubramanian

>> Yeah. To me, first I would say agentic AI in my mind is going to be one of the biggest technology transformation in the history of tech. And to that end, at AWS, we are building for the foundation for the future of billions of agents operating in the next few years. And to me, that is an understatement. There are going to be a lot more. So now in such a world, what does it mean for us to prepare? And that is really one of the reasons why we are investing in Frontier agents, because we want to reinvent how software development is done in the world. And if you see the history of what is going on, and when I say history, this is literally past 18 months. And this based on software development assistance and-

John Furrier

>> It's the opening app.

Swami Sivasubramanian

>> Agents.

John Furrier

>> It's the opening act of gen AI.

Swami Sivasubramanian

>> Exactly. Right. But so much focus have been only around writing code. But if you see what is happening in a true enterprise or anywhere, it is like 20 to 30% of what a developer does. They end up spending a lot of time on DevOps and dealing with on call and SRE work. They end up on security and upgrades and various other things and pen testing and whatnot. So to some extent, we did not do enough to give them the freedom to ideate and create. And so we asked ourselves saying, like let's reinvent this whole discipline and come up with virtual agentic teammates. So initially, even moving beyond vibe coding, first we started with creating Kiro, which is our agentic IDE, where we said, no longer do you just need to do vibe code. Let's actually use agents to create specifications on what you're trying to accomplish and then build ideas, I mean, or design and then create code and write tests. So this led to this whole transformation where you saw Matt talked about how we had a team of six engineers built this whole inference thing, what could have taken us like 30 to 40 people for one year. We got it done in like 76 days. That is like an example of suddenly we just changed the game. But then based on that experience, we learned something. We said, "You know what? Why actually even in many cases you need humans to steer? Why can't we have agentic teammates come in?" And that led to Frontier agents with key autonomously planning and DevOps agent is taking care of uncommon product to SRE. And then you have actually with security agent doing pen testing and design and everything. So this is going to be a step change in productivity where you're not talking about 20 to 30% gain. We are talking about 5X to 10X.

John Furrier

>> I love Matt's line. I want to get your thoughts because you just kind of hinted at it. "Agents stop being toys, they become teammates." One of the things in the Kiro pre-interviews I did, the thing that jumped out at me was it learns your preferences, like-

Swami Sivasubramanian

>> That's right.

John Furrier

>> How you document certain, the nuances of teams.

Swami Sivasubramanian

>> That's right.

John Furrier

>> Certain teams have particulars-

Swami Sivasubramanian

>> Yeah.

John Furrier

>> That they like certain things the way they are.

Swami Sivasubramanian

>> They're very opinionated about it, so.

John Furrier

>> Of course it's easy to yell at an agent. There's no hostile work environment there. But I mean, talk about why the teammates thing is so important because that makes it personal. That makes it fun.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> And makes it productive.

Swami Sivasubramanian

>> That's right. To some extent, the key thing with these agents showing up as teammates is because then they learn your key code bases, they learn your preferences and then they have built in right memory so that they even know, for instance, like, hey, this team actually ends up coding up with this style even as what is the preference on how they like to do code review to what are the ways they like to upgrade particular thing? What is their favorite databases to whatnot and then it starts building context and brain end to end so that it doesn't ask dumb question. The problem right now with most AI assistants are that they always act more like an intern instead of a very tenured teammate within the company. So that means every day is like day one and you keep training the intern.

John Furrier

>> Yeah. In DevOps, they talk about day two operations. Your DevOps Frontier agent essentially is a day two operations engineer-

Swami Sivasubramanian

>> That's exactly right.

John Furrier

>> That's on call.

Swami Sivasubramanian

>> That's exactly right. And it actually is not just an on call engineer, but it also has another, I call it the custom of an SRE agent which does proactive SRE work. And so when you think about that one, them showing up in teams Slack and saying like when I used to get paged when I was a principal engineer, I still get paged when incidents happen, but when I was on call as an engineer, if I wake up at 2:00 AM in the morning, earlier I used to be the one debugging the metrics and logs and everything else. And we still did. But now with DevOps agent, we have the possibility of headfirst engaging at the same time so that by the time I log onto the computer, now it can notify me on Slack saying like, I did all the background work, so here is most likely the root cause.

John Furrier

>> It's like a chef. You have a sous chef. They chop the vegetables and everything's ready for you. They're ready to go.

Swami Sivasubramanian

>> That's right.

John Furrier

>> You mentioned earlier that you guys really had a vision to transform software development lifecycle.

Swami Sivasubramanian

>> That's right.

John Furrier

>> And this is one of the motivation. You and I have talked about this in the past. Now it's on stage in the keynotes. This applies not only to building apps, but to also doing other things. I saw the display on Monday, I think it was, where you blew up a legacy system.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> I don't know which one it was, but migrations is in the agenda. I mean, like migrations, what the hell? I mean, that's top line story, but the point is that's still software.

Swami Sivasubramanian

>> That's right.

John Furrier

>> So that again, so the software reinvention could apply to migrations, it could apply to code reviews, it could apply to everything.

Swami Sivasubramanian

>> Everything. To me, migrations arguably is one of the, it is one of the most tedious things that is really slowing down every company to reinvent them. This modernization being stuck in mind frames or like expensive.net licenses or what's going on at VMware and so forth, let alone even things like, You know what? I need to upgrade from V1 of this library to V2 or we need to move from pull to Java and whatnot. So this is why when I started this new organization I said, "You know what? Let's make modernization cool again by making it agentic. Let's actually change the game by doing an order of magnitude faster." So that's why mainframe migration no longer is like this three to four year thing. It is like now done in months and for the mainframe part. And same is true for this custom transformation. So when I was talking to Julia and then saying like, "This is really profound transformation that you need to showcase in a way." And Julia came up with this brilliant idea of, "Let's blow our technical debt." And I'm like, "Yes, that's what we are doing." She said, "No, no, no. You got to actually literally blow up." And I was like, "Holy shit, that was fun."

John Furrier

>> That was fun. It was awesome. It's very showy, but it makes the point.

Swami Sivasubramanian

>> Exactly.

John Furrier

>> Most migrations have been months.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> And it's tedious, it's grinding.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> And now you can literally clone and create digital twins. There's now a big movement from some of the advances that you guys announced where people are looking at code bugs, fixing them fast. We're going to get a bunch of folks here on theCUBE. I just saw a company get funding. They're doing a simulation and essentially pre-vetting code before it even goes into GitHub. So you're starting to see kind of new techniques.

Swami Sivasubramanian

>> That's right.

John Furrier

>> Are there things that you see that are going to come out of this? And one other observation I'll share with you and get your thoughts on is there's more data coming off these agents too because there's telemetry and there's also a vector embed, has more data than just Word. So now that's more storage. So how do you view that and what does the infrastructure need to do? Obviously Trainium's got advances, tokens per the watt as a metric. There's more data coming in.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> And the budgets aren't necessarily going up as much.

Swami Sivasubramanian

>> Yeah. I think-

John Furrier

>> So I mean pre-agentic the budget, the data growth was like 30 to 40% CAGR.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> Now it's going to probably be hundreds.

Swami Sivasubramanian

>> Yeah. So that's a great question, especially in a world where more and more agentic workflows are going to happen. Many thinking problems are becoming compute and data problems. So that means fundamentally you continue to drive down efficiencies on both the compute and how we store data and how these agents are getting built. So that's why things like investment in Trainium, let alone what we had done with the vector industry with S3 vectors and open search are game changers, along with now with things like AgentCore, how we are able to build these short-term and long-term memories and episodic memories.

John Furrier

>> Yeah.

Swami Sivasubramanian

>> So now think of what we have done with like Transform custom. When we are upgrading 100,000 repos, then suddenly we can learn what worked in one repo and what failed and then we can apply it in a different context on the, let's say the 1,000 and 1st repository. So then that means this agent is continuously learning. So it gets better fast, really quick after first time.

John Furrier

>> And that's been a Frontier agent versus a-

Swami Sivasubramanian

>> That's exactly-

John Furrier

>> A regular agent.

Swami Sivasubramanian

>> That's exactly right.

John Furrier

>> Almost like LLMs before they had Frontier capabilities-

Swami Sivasubramanian

>> That's right.

John Furrier

>> Which was multi-step reasoning, et cetera, et cetera.

Swami Sivasubramanian

>> That's exactly right.

John Furrier

>> I have to ask you on the intelligence side. If you look at, say, AgentCore, you've got memory identity gateway, browser tools, observability, drift detectors, policy enforcement, natural versus natural language. It's all in there, but also Strands made it back to the main stage again.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> So last time we talked, you were bragging about Strands, being an internal project and proud of it. That has become quite the product.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> What is Strands now versus when it evolved?

Swami Sivasubramanian

>> I know.

John Furrier

>> Because it's come into, I've interviewed Clare Liguori in Seattle. She was proud mom of that product.

Swami Sivasubramanian

>> That's right.

John Furrier

>> But where is Strands in the scheme of the agents? What's the role of Strands?

Swami Sivasubramanian

>> I think Strands is one of the amazing invention. Clare is the big pioneers for it, which really came from an era where we say, hey, in this world where LLMs have this amazing reasoning capabilities, how we build agents can be done remarkably in a simple manner. So she and couple of, literally two others built this framework to build things like Kiro. And then we saw so much that where we were able to delete thousands of lines of code and get better accuracy with this framework. And I still remember she actually messaged me, I think this was sometime in summer at seven o'clock in the morning saying, "This is so impactful that we think it is better." And she thought it was better that we open sourced it so that all AWS customers, let alone everybody in the industry could actually build better agents and it'll move the industry forward. She said, "So we should open source it." So literally, I think three hours later, Clare and I and Marc Brooker and few of us met and we decided, they wrote like a one pager on pros and cons and we said, "Okay." After that meeting, we decided we should." And that evening I met with Matt for like 15 minutes and Matt said, "Okay, let's do it." And then we open source the day after. That is literally how fast we code. Then faster to now, it is by default, people are able to leverage trans to build agents in such a breathtaking phase like what you saw with Cox Automotive to what you're seeing with several others and what you saw with like Blue Origin and whatnot. So now what we are seeing is essentially that who can build agents is changing so fast. No longer do you need all this complex orchestration logic. And it is taking off at like five million downloads in such a short time. And the pace at which this is getting used is also accelerating month over month. So it's so many community contributions and whatnots.

John Furrier

>> All right. So I have to ask you, because I asked a bunch of other folks this question too, this re:Invent seems to be one of those moments for AWS where the ground's shifting, but the game is still the same. You got infrastructure as a service, platform as a service, and now AI native apps. We interviewed Esquire on container native.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> So you got now the Kubernetes piece, cloud native and IA native coming together. What do you expect to change after this re:Invent? What do you expect to see? People are already enthusiastic about agents.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> Is it going to be more confidence? Is going to be more unlocking the enterprise? What do you see as the next chapter that will come out of this re:Invent?

Swami Sivasubramanian

>> Well, I tended to say you should wait till next year re:Invent to see. But I would say in many ways, what changes. While it's a very important question, a more important question is going to be what's not going to change primarily. And for that, people are going to actually want agents that they can rely on and they can actually find it trustworthy. And that it solves them problem that takes care of undifferentiated every lifting. While we are starting to see, and many of the innovations that I touched on today and Matt touched on, is laying the foundations. You are going to see who can build an agent is going to change fundamentally. That means every app nowadays has a compute storage database and some AI backed in. But every app by next year is going to be agentic. And that is going to be a given in my mind.

John Furrier

>> That's the next era.

Swami Sivasubramanian

>> And that means the world is about to shift where it is going to be completely agentic driven and it is going to be the new normal and that's what we are empowering.

John Furrier

>> Swami, great to have you on. Again, congratulations on great keynote. I know you worked hard as a team. Love the speed of the transformation, love the rethinking. And Matt used an example with Reddit.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> Domain expertise is huge in these agents.

Swami Sivasubramanian

>> That's right.

John Furrier

>> So speak quickly about the domain specific nature of why that's important, because this is where you got the MCP, AWA, you got the multiple frameworks, including Strands. This choice now to code anything up, whether it's Nova or another model in Bedrock, SageMaker has now bidirectional support for audio through Deepgram. We interviewed them. A lot of other innovators are coming into the system.

Swami Sivasubramanian

>> Yeah.

John Furrier

>> And it's a systems architecture.

Swami Sivasubramanian

>> Yeah. I mean, domain expertise is super important in this new world. The nice thing in this agentic world is that you can bring in your domain expertise in multiple layers of the stack. Yes, you can bring it in as APIs that you bring in through MCP as you pointed out and you can integrate it directly with even out of the box tools like Quick Suite and bring in your own MCP too. You can also enable it with developers and build it through AgentCore. And of course you can go then one layer deeper with the model level and customize it using tools like NovaForge and everything. It is because at the end of the day, the way I think about domain expertise is every business has to think about what is the core underlying value and asset that they have and how they would like to leverage it and integrate with everybody else. And for many, it is about like, how do I convert my SaaS and API business and expose it with other agents? So that's why they're exposing it as MCP and A2A.

John Furrier

>> Yes.

Swami Sivasubramanian

>> And for others, it is about like, I would love to build my own Frontier model. And that is where things-

John Furrier

>> Yeah. Generic reasoning doesn't work in these domains.

Swami Sivasubramanian

>> That's right.

John Furrier

>> Matt said, I'll end the quote here, "Generic tokens are useless unless they know your business."

Swami Sivasubramanian

>> That's right.

John Furrier

>> That's where agents shine.

Swami Sivasubramanian

>> That's right.

John Furrier

>> You agree with that?

Swami Sivasubramanian

>> That's right.

John Furrier

>> Swami, great to have you on. I know you're super busy. Thank you for-

Swami Sivasubramanian

>> Thank you.

John Furrier

>> Spending that time to come into theCUBE and our annual tradition.

Swami Sivasubramanian

>> All right.

John Furrier

>> I missed you in Seattle. Thanks for spending the time and sharing your insights.

Swami Sivasubramanian

>> No, my pleasure. Thanks for having me and hopefully we get to do it again.

John Furrier

>> We'll be here next year. If not, our agents will be here working on our behalf. I'm John Furrier with theCUBE. Thanks for watching.