Varun Jasti of AWS explains why real-time data—not better models—is the true unlock for enterprise AI.
Most enterprises don't need to build AI models from scratch—they need to put AI to work. That requires a data foundation that is real-time, reliable, and ready to serve intelligent systems at scale.
In this episode, Varun Jasti, Solution Architect at AWS, joins host Joseph Morais to explore how AWS and Confluent are co-architecting the infrastructure layer for agentic AI. Varun draws a direct parallel between multi-agent system design and microservice architecture—explaining why Apache Kafka® on Confluent Cloud is the natural backbone for agent-to-agent coordination, stateful memory, and scalable provisioning.
You'll Learn:
About the Guest:
Varum Jastim is a Solution Architect at AWS, and an AI/ML enthusiast that likes building on the cloud to architect solutions for AWS Partners who need to adhere to federal compliances and regulations. His areas of interest are Computer Vision and Generative AI.
Guest Highlight:
"The average company is not an AI lab. Their job is not to say, 'How do I create a great AI model?' Their job is, how do I use AI to create business value? And I believe that's the core mission of AWS and Confluent to get together and figure out how we make AI accessible to the customer."
Chapters:
[01:23] Why AI Feels Different
[04:07] What Makes AI So Provocative
[07:46] Amazon Bedrock AgentCore Explained
[09:20] Real-Time Context Engine
[11:57] Scaling Multi-Agent Systems
[16:01] Where to Run Agents
[18:36] Security and Guardrails
[26:24] Global GenAI Architecture
[29:21] What's Next at AWS
Dive Deeper into Data Streaming:
Get Connected:
Our Sponsor:
Your data shouldn’t be a problem to manage. It should be your superpower. The Confluent data streaming platform transforms organizations with trustworthy, real-time data that seamlessly spans your entire environment and powers innovation across every use case. Create smarter, deploy faster, and maximize efficiency with a true data streaming platform from the pioneers in data streaming. Learn more at confluent.io.
[music]
0:00:40.9 Joseph Morais: Hello everyone. Welcome to Life Is But a Stream. As always, I'm your host, Joseph Morais. Today's episode is a very special one. It's the first of our Cloud Connect series where we spotlight how Confluent and our cloud partners are shaping the future of enterprise innovation through real-time data streaming. Each of these episodes features conversations on how organizations are modernizing infrastructure, accelerating AI adoption, and building intelligent connected systems in the cloud. Today I'm talking to a Solutions Architect from AWS on how Confluent and Amazon technologies can create differentiated AI outcomes for enterprises of any size. Let's get started.
0:00:46.7 Joseph Morais: Varun, thank you so much for joining me on the show today. So Varun is from our strategic excellent partner Amazon Web Services. So Varun, why don't you tell the audience a bit more about yourself and what you do at AWS.
0:00:57.7 Varun Jasti: Thank you, Joseph, for having me on this podcast first and foremost. My name is Varun Jasti. I'm a Partner Solutions Architect at AWS and I work with the Confluent partnership to figure out how we can co-innovate for our customers to help solve their data streaming needs.
0:01:13.8 Joseph Morais: Today we're, as I mentioned in the introduction, we're talking about agentic AI and the outcomes that are possible using a combination of Confluent and AWS technologies. What's really fascinating to me, Varun, is how fast everything is moving. I'm a child of the 80s and I know that there's been these different progressions, whether it's the internet and then mobile became big, but things with AI are moving really fast. Do you have any thoughts on that?
0:01:40.3 Varun Jasti: It's really interesting about how everyone kind of knows about AI. There are a lot of people trying to use AI and at enterprise scale. I actually read this interesting blog slash memo that was written by Howard Marks. He's one of the Co-Chairmen at Oaktree Capital Management, which is one of the big hedge funds. And he said that enterprises are at almost 75 to 80 percent of AI adoption. So what that means is we have 75 to 80 percent of enterprises trying to adopt AI or figuring out how to adopt AI and how to use AI. That causes a whole bunch of interesting questions and scales because the average company is not an AI lab. Their job is not to say, "Hey, how do I create a great AI model?" And another thing that's just as important is their job is also not to say, "Hey, how do I build all this AI infrastructure?" Their job is how do I use AI to create the business value, and I believe that's the core mission of AWS and Confluent. To get together and figure out how do we make AI accessible to the customer, where on the AWS side we figure out how to provide the AI infrastructure and on the Confluent side, you guys do a great job of figuring out how to bring data accessible to the customers, because AI is going to be only as useful as the data you feed it. And that's where Confluent plays a very valuable role as our partner in being able to give our customers the ability to access data when, how, and also proper manner.
0:03:11.2 Joseph Morais: Yeah, I couldn't agree with you more. There's actually a picture of me from 2023 at Current in San Jose, and I'm holding up a sign that says, "Your AI is only as good as your data." So I was early on that trend. Something you mentioned to me, or you mentioned that I think is really fascinating is really the broad curiosity that we see across enterprises and across our own individual relationships as it pertains to AI. So as you mentioned, computers became more mainstream in the 80s, more accessible, but I would argue that they weren't fully adopted. Every house didn't have a computer till probably closer to the early 2000s and making that gap even larger. But it's not just technologists or IT people that are talking about AI in 2024, it's everyone. And I can't really think of anything like that, except maybe the internet. But that definitely was a way slower burn than this. What do you think...
0:04:07.1 Varun Jasti: Yeah, I remember that.
0:04:09.5 Joseph Morais: Yeah, I do remember that. What do you think is so provocative about artificial intelligence that everyone, regardless of enterprise and... And I would take a step further, I would argue most enterprises aren't even technology companies, let alone AI companies, and they're still struggling to do things like data streaming and even microservices. Why do you think this topic is so provocative and every CEO, it's on the top of their mind?
0:04:29.8 Varun Jasti: I think what's made AI possible even is how hardware has become a lot more commoditized. It's a lot cheaper and scalable to acquire hardware. It's a lot easier to get bandwidth for the internet. In order for an AI to work and give you fast, relevant answers, you also need good bandwidth. And you also need to have good memory. These are all the things an AI needs and requires, that these are all things that have been accessible. And since they're accessible to the common man today, when they see an LLM model return results such as, "Hey, can you write a history paper for me?" or "Hey, can you write a piece of code for me?", it's able to do it really fast. And what this does is accelerates productivity. Similar to how I said, the average enterprise that's out there, they're not an AI lab. They're not necessarily a pure technology company. What technology does is it's given all these companies the power to accelerate their business goals. It's similar to a database. When databases came out, they revolutionized how a company can access data and they can maintain consistency for that data.
0:05:39.7 Varun Jasti: It's like scales of innovation. AI is something where you can put everything together in one spot and accelerate productivity at each and every aspect of the company. Because it's not that AI is just affecting one piece of the company, it can affect your entire organization, whether it's for your internal organizational productivity use cases or to help make the customer journey easy on the external-facing use case. And how an AI can give your customers answers a lot faster than let's say a human agent can when they have to wait and call, especially in support.
0:06:14.5 Joseph Morais: Sure.
0:06:14.9 Varun Jasti: When especially back in late January when we had the winter weather on the East Coast and you're trying to call up American Airlines and you wait for an agent, it's a lot slower process than me just asking one of their chatbots a question and then they spit out an answer.
0:06:32.8 Joseph Morais: Yeah, absolutely. I think you nailed it. The accessibility is huge because it's not just demonstrations. People aren't just watching YouTube videos and seeing somebody use this thing and they can't access it themselves. They can literally go to a website and start interacting with it in that moment. And I would also add to the fact that the syntax... So let's say you were interested in databases, as you mentioned, great example. There's still a barrier to databases. You still have to understand the query language, you still have to understand SQL, you still have to access the construct of a database, which is non-trivial. But the fact that these APIs or these chatbots are accessible right over a web browser and they're in natural language, you can interact with them the way you interact with anything in the world, is really kind of mind-blowing. So I had someone on the show who works for Cursor and he mentioned that coding's a bug. At the end of the day, what you want is the computer to do what you want. And we had to, or at least up until very recently, we had to interact with it in a very specific syntax so it understood our instruction set.But this whole natural language layer over top of it is really just incredible. And it's not just limited, of course, to coding, but the way we interact with tooling, the way we interact with AWS, the APIs, the CLIs, it's really changed everything.
0:07:48.8 Joseph Morais: But kind of fitting into that theme of accessibility, let's talk about some of the tools that can make agentic AI accessible to any enterprise. And let's start with AgentCore. So for the uninitiated, there is a subservice under the umbrella of Bedrock called AgentCore. How does AgentCore differ from some of the other services that you could potentially run agents on in AWS?
0:08:12.1 Varun Jasti: The barrier to entry is a lot less in terms of the end user or developer doesn't need to learn a particular syntax or query or language in order to get a system to do what it wants. AgentCore follows a similar process, where it makes it very easy for you to provision production-level infrastructure for your AI agents to run without having to go through the manual process of having to spin up Kubernetes clusters, for example, and putting your agents on top of those clusters and then making sure the clusters are running appropriately and you're not waking up at 2:00 AM getting pinged because there's a patch that needs to be fixed. The maintenance of it. That is something where AWS will provide and take that responsibility as part of a shared responsibility model to provide that production infrastructure for your agents by giving it the proper tools, memory, syntax, and also session isolation. That's where Confluent also comes in, as you guys also work very well with our AgentCore product in order to give it the proper data in time. Because an agent is going to be only as good as the context that's given. And that's where you come in.
0:09:26.8 Joseph Morais: That's true. Well, okay, and then this is a perfect opportunity for me to talk about a service that we launched late last year during Current in New Orleans, which was an absolute blast, and a plug for Current London coming up in May 2026. We launched something called the Real-Time Context Engine. So followers of this show, you know that we're in the world of data streaming, so we always talk about Kafka and things like Flink, stream processing and data streaming in general. But what we have is a new feature, a part of Confluent Cloud, which is our fully managed offering available on AWS, and it's called the Real-Time Context Engine. And what it is, is it's delivered as an MCP server. So again, that Model Control Protocol that has gained so much popularity around providing tooling to agents. And what it does is it presents everything that's in Confluent Cloud in that MCP decoupled layer. So suddenly your agents know about your clusters, it knows about context, which is a very big theme of 2026. What does the agent need to know to satisfy the request? Because if you just ask an agent a question and you don't feed it the proper data about your business, it's just gonna respond based on what it was trained on. But it wasn't trained on the encompassing data of your enterprise, and that's where the context engine comes in. So you could have an agent running on AgentCore and then utilizing the Real-Time Context Engine to have it know exactly what you need that's relevant to your business.
0:10:55.8 Joseph Morais: And that really kind of comes down to a lot of different parts. Because very rarely in an enterprise is all of your data in one place. If you're a large enterprise and you say you have one database, I challenge you on that. Realistically, the data that your agent needs is spread throughout your estate, and that could be exclusively on AWS in a single region, it could be across multiple regions, it could be a combination on-prem, or you could be just evaluating your journey to the cloud. And you need to be able to have a way to tie all that data together. And that's really where Confluent really shines. So we have products that run on-premise and in the cloud, of course, and then we have these number of connectors that allow you to integrate your existing data systems. So think about legacy databases, mainframes, HDFS, anything you can think of, and put it into a single substrate so that all of your data lives in one place, Confluent Cloud on AWS, and then is served natively to those agents through the Real-Time Context Engine. So AWS handling kind of all of the pain points of running the agents, where Confluent can handle all of the pain points of feeding the data to those agents.
0:12:07.6 Joseph Morais: Now, I'm curious about AgentCore. It's pretty trivial to get a single agent running. Production in enterprise production, we're talking about scaling. So are there any particular patterns that you see AWS customers using for scaling from a single agent to many agents and more importantly those agents working together?
0:12:25.0 Varun Jasti: AgentCore does a great job of helping you run a multi-agent autonomous system. But before we go too much into what AgentCore does and how you're running this multi-agent autonomous system, I think it's important for enterprises to kind of understand what they're building. When it comes to, "Hey, I'm building an autonomous AI system," sounds cool in its context and a pretty cool thing to build. You kind of have to understand the fundamental task that each agent is accomplishing. And when you look at how agents and multi-agent systems are designed, it's really no different from how a microservices architecture is designed where, similar to a microservice, where each microservice has its data that's streaming in, that's being shared between the other microservices. Each microservice has its own storage, it has its own memory, it has its own database architecture, similar thing like that. An agent's gonna have its own data that gets streamed to it. An agent is gonna have its own handoff and receive mechanism where inter-agent communication is there. And that's where you have protocols such as A2A that Google came out with. And you have these mechanisms where agents are pretty much, like I said, running similar to these microservices.
0:13:40.5 Varun Jasti: And when you're interconnecting, let's say, a system with thousands of agents together, you need an infrastructure service that you can provision out of the box with a simple Docker deployment, actually with a Dockerfile deployment, and that's managed for you. That will save you lots of time. And that's where AgentCore comes into the picture here. And I've had customers that have been using AgentCore come up to me and said, "Hey, man, thank you so much for AWS for coming out with this service. Because now I don't have to worry about managing the underlying infrastructure of all these agent systems."
0:14:19.2 Joseph Morais: Sure.
0:14:19.5 Varun Jasti: "And since I don't have to manage this underlying infrastructure, I can actually focus on the more valuable aspects of my job and in terms of how we can use these agents to complete our business goals." That's kind of where AgentCore comes into place. And then another thing AgentCore provides is an ability to also natively interact with MCP and tools. Because you mentioned that Confluent's Real-Time Context Engine, it uses MCP. And that's kind of where AgentCore, and provides you a natural gateway mechanism to connect to your MCP tooling. And I feel that's also the most important part because an agent is only going to be as good as the tools you provide it. And that's where AgentCore gives you the infrastructure to connect to MCP.
0:15:06.8 Joseph Morais: Yeah. Well, I love your comparison to microservices. It's one I make pretty often because the worst thing you could possibly build is a monolithic agent, right?
0:15:15.1 Varun Jasti: Yes.
0:15:15.6 Joseph Morais: Where you put everything in this gigantic prompt, and it really is not good at any one thing. And there's also things like context windows. If you try to do everything in a single agent, you're gonna run out of a context window, you're gonna deal with compression... Compaction, and that's gonna make it slow and it's not gonna work very well. So by having, chunking up your agents into specific functions, you can of course shrink those context windows. But also for troubleshooting, it makes it a lot easier when you have everything chunked up to send an input and look at the output at these little segments of the process, as opposed to, "Here's a giant prompt, here's all of my data, do something for me," and what the black box that you fed it into outputs something and you have no way to reconcile how it came up with any of that. So I think it's very pragmatic to build these superpowered microservices which we're now calling AI agents.
0:16:12.0 Joseph Morais: Now, speaking of running agents, there's a number of different ways to do that in AWS, including in Confluent Cloud. So another feature we launched last year was called Streaming Agents, where you can actually run agents right inside of our Flink offering inside of Confluent Cloud. So literally, we're managing the runtimes for you in that scenario. So that's another place in addition to AgentCore. But there's also places like EKS and ECS. So if I'm a prospective or current AWS customer, what kind of decision tree am I going through to make a choice? Where am I running my agents? Should I be running them exclusively in one place, or is there perhaps a known pattern for where you want to run these agents in AgentCore versus EKS versus ECS versus Streaming Agents in Confluent Cloud?
0:16:53.7 Varun Jasti: It depends on a few key factors, but I would say the most important is how much control do I want of the infrastructure that I'm deploying.
0:17:03.4 Joseph Morais: Interesting.
0:17:04.5 Varun Jasti: Because if you were... Don't even have to go as far as EKS. You can even deploy them on EC2s if you want. But that wouldn't be the most efficient in terms of... Because that's more management overload if you're trying to manage them on an EC2. The idea here is that how much control do I want and also how much control do I need and also what are my compliance mechanisms that my organization has to follow. At Amazon, we provide the managed infrastructure, with AgentCore you don't have to worry about managing, like I previously mentioned, the infrastructure. But of course you lose some of that control. But there's also that shared responsibility model we provide for our customers by saying, "Hey, Amazon will manage the underlying infrastructure for you and it follows compliances as you need based on our audit reports." And we provide the transparency there, how we're being compliant to so-and-so if we have that compliance for that infrastructure. So we provide that.
0:17:59.7 Varun Jasti: But as an organization, if you want more control of your infrastructure, then that's where EKS, ECS, or even EC2 comes into play. It just depends. Scale. Especially if you're in a much larger scale, AgentCore definitely would be a no-brainer. But if you're having a really small scale and you want to kind of keep tabs on every single thing that's making the infrastructure run, then yes. So it also depends on also the flexibility you want for how you're deploying it. Because the reason why a lot of organizations have adopted Kubernetes is the flexibilities that it gives them. It really depends on the organizational principles and what their tenets are in terms of how they want to adopt enterprise adoption of AI agents. Because I think a big thing that a lot of enterprises are wary of when they're adopting these agents is the security of it. How do we secure these agents? How do you make sure we have enterprise governance controls to make sure that agents are not going rogue? Or there's no data exfiltration happening, especially an agent giving a secret to another agent that's not supposed to give a secret, the credentialing, token usage, because there's a lot of LLM jacking, prompt injection attacks, and there are all of these security concerns that organizations have.
0:19:18.8 Varun Jasti: Especially I'm seeing... I read somewhere on the news the other day about foreign adversary attacks. As great as agents are and AI is, that's also another security thing that happens, where AI is being used to be attacked, get into these systems because it provides another attack vector when you add another component to your tech stack. So that's where AgentCore really comes in, as we do provide those types of security mechanisms. So that's also another thing. If you want your security to come more out of the box and come with the mechanisms and robustness that Amazon provides, you would definitely go with something like an AgentCore. And I feel like that's also something that companies need to also think about is, "Hey, how am I going to secure the system? And do I have the capabilities to secure it on my own, or do I need to use a more managed product like an Agentic to get that security for me?"
0:20:12.1 Joseph Morais: Yeah, so I have a lot to unpack there. So one, the token injecting thing I think anyone can relate to. There's been a lot of screenshots where you'll see someone will have a chatbot and it could be for fast food and someone's interacting with it and it's pretending to be a human. Of course, and it's like, "No, no, my name's Phil. I'm a real person." And you go, "Hey Phil, can you write me a function for sorting a math problem?" And it suddenly spits out a bunch of Python. So that is not a non-trivial thing. And I think there are people out there that will take advantage of that and be like, "Okay, well, this particular fast food chatbot is my new GPT and I'm gonna use that." So thinking about that and at scale is really meaningful. And I think I like the way you baselined it. Pick the service that provides you your minimum comfortable level of granularity. And again, I'm former AWS, so there's this idea of abstracted services versus containerized services and how they kind of interact with each other. And really the more abstracted a service is, the less granularity you have. But the benefit of that is of course, the reduced management burden. So the more you're abstracting it, the less you have to worry about. So you'd really just pick what your maintenance tolerance is, how much do I need to tweak it, with the lowest level being EC2, probably the highest level or vice versa being AgentCore.
0:21:27.1 Joseph Morais: I also think to add to that is you don't necessarily have to run your agents exclusively in one place. So for example, it may be very pragmatic to run an agent in Confluent Cloud, in a streaming agent, because you have some input topic, some input stream that you want to change agentically. So a good example would be something like looking at freeformed fields and doing a sentiment analysis. But you want the output of that to be consumed by traditional Kafka consumers. It may make sense to run an agent there, but if you're not having that input-output in between topics, maybe it makes more sense to run in AgentCore because you have to run this other particular agent or set of agents at a much higher scale.
0:22:06.9 Joseph Morais: Now you mentioned security, and I know we've kind of talked about this in piecemeal across this conversation, but can you provide what I think would be maybe an exhaustive high-level list of what agents really need to run in production? They need data, they need session isolation. What are the other things that really come top of mind, hat would come top of mind for any enterprise regardless of their size?
0:22:26.8 Varun Jasti: Memory, right? An agent needs to be able to have context of the conversation to be able to better help the end user that's using the agentic product. That's another piece of it. And then double-clicking more on session isolation, that's also very important because not only does it maintain the session state and you get the stateful sessions piece of it, there's also the fact that it's being run with the microVM, and that provides you the privacy, reliability, and security that your agent would need. And also as an enterprise, you would want to keep to make sure that an agent is operating only there and you don't have any cross-contamination between other agents running and flowing in. Because at the end of the day, unlike microservice architecture, a multi-agent agentic architecture is... When you run a microservice, it's very defined and it is a deterministic outcome, because you have code that's written to do a specific task. Now an agentic AI system where we have each agent, it's kind of like a microservice with a brain. And the brain part is very key there because it's non-deterministic. It starts doing some thinking on its own.
0:23:38.4 Varun Jasti: So guardrails, you need to be able to set up proper guardrails. That's kind of where security goes really in-depth, network access, and we can go all day talking about security and of course, that's not the purpose of why we're here. But it's also very important to say like, hey, security is very important because that's kind of how hackers are getting into these enterprise systems. And AWS does provide that security of the underlying infrastructure that we provide. As I say, shared responsibility model, the customer's responsible for how they're securing the system that they're putting on top of the infrastructure. Confluent, I would say also, if you can bring that into the conversation here is, provides a great job of being able to secure the data, especially on Confluent Cloud, especially when you're streaming from these Kafka clusters. That's another part. And also, you brought up a good point of where you run the agents. And that's kind of where if you have also other integration points, you kind of want to minimize the hops that an agent takes. Not only for the performance reasons but also for security reasons because you have to reduce that attack surface. It kind of all blends together. And honestly, just to say, we would love for customers to also be able to run their agents on Confluent Cloud 'cause it's built on AWS. So we win either way.
0:24:59.1 Joseph Morais: Right. Why argue over what flavor of ice cream when AWS makes all the milk? I've said that many times. I like the way you put that down... You kind of put that together. Agents are the opposite of idempotent. You can send the same input 10 times, you're gonna get 10 different outputs, and it's because it's so non-deterministic. So I would argue that it is more important to secure these agents than it is microservices simply because of that nature. And I appreciate the nod to Confluent Cloud and the data streaming platform. Varun mentioned the... Because we can source data from everywhere, which I mentioned earlier, but we can also stream process. So as opposed to giving an agent direct access to a database where it may have all of the access, including things like PII, protected information, healthcare data, very sensitive information, we don't have to present that to our agents. We can integrate all that data into data streams and then use stream processing to stitch different events together, but then also filter things out. Like you want to hide a Social Security number, we can do that. You want to hide keys, important hashed fields, we can do all that and only provide the data to the agent that it needs, and then all the other topics that it doesn't have access to, it just wouldn't have access to.
0:25:29.6 Joseph Morais: So you can secure it in a different way by sanitizing that data and presenting what the agent needs to make that decision without potentially giving it anything to leak, which is a whole 'nother level of security and very attractive for most enterprises. So beyond just the integration, you also get assurances that the data is of high quality through our governance systems and our services, but also in the way that we can deliver explicitly contextualized data, however we want to present it to our agents.
0:26:39.2 Joseph Morais: Now, I have a networking background. I'm a bit of a network geek, and I kind of geek out when we get to use services like Transit Gateway and Direct Connect, and customers have this, especially large enterprises have this hybrid relationship with AWS. Is there an appetite that you've seen with customers for kind of a global gen AI architecture, or things kind of starting still in single regions or people thinking bigger at this point?
0:27:06.6 Varun Jasti: Honestly, depends on the scale of the enterprise.
0:27:09.0 Joseph Morais: Sure.
0:27:11.5 Varun Jasti: But we're definitely seeing a global approach to gen AI infrastructure because gen AI is something that's been taken over globally. You need to be able to provide the service at a global scale. Another thing that is very important for an organization is that for their systems to be able to have that resiliency and also have redundancy built in as needed to be able to work as an always-on system depending on the RTO/RPO objectives. Another thing also is languages. Different languages. And in order to reach a global scale, you need to be able to provide them in a different global language. And you also want to make sure that you're able to run close to the customer and also on capacity to be able to get the capacity you need. And in order to get all these things, because especially there's a GPU shortage and that's why Nvidia is the multi-trillion dollar company that they are today, is that you need to be able to get capacity where you can, and AWS does a great job of providing that capacity at a global scale. So due to that reason, and also for business reasons, I do see it going towards a more global scale. And I'm seeing it a lot. And when you want to get capacity on demand, the best place to go is the cloud, and the best cloud there is is AWS, right?
0:28:30.4 Joseph Morais: So I like that. So again, obviously I was kind of leading a little, but that's where Confluent and AWS together can really help you. You have these great networking technologies, like I mentioned, TGW, Transit Gateway, Direct Connect, Private Link, that help you bridge your on-premise data and across regions and everything. But then Confluent enables and works with all of those. We have clusters that enable Private Link for private communications only in your VPC. You can run clusters over Transit Gateway and allow your clusters to work across the entire all the estate of AWS and have a hub-and-spoke model. You can run multi-region, so have clusters in every region sharing that data back and forth through a technology that is unique to Confluent called Cluster Linking. It is the best data streaming replication technology in the industry, and I will die on a hill on that one. So between these network technologies from AWS and these data-level technologies that can help with replication and integration, you can really build a global data mesh so that you can take advantage of GPUs in any region regardless of where they are. And that's something I hadn't even really thought of.
0:29:35.7 Joseph Morais: This has just been a really fascinating conversation, Varun. Before I let you go, I'd love for you to kind of leave the audience with things to get excited for. So we're almost at the end of Q3 in 2026. Are there any recently launched features, services, maybe things that are in preview or about to come out that you can talk about that people should be excited for when they're thinking about agentic outcomes on AWS?
0:29:58.1 Varun Jasti: Yeah. So in re: Invent, we launched a DevOps agent. That's been a great unveiling for us. It's in preview now. It's about to go to GA soon. And that's a great tool to use to have an always-on engineer, site reliability engineer, that can automate incident responses and give that detailed mitigation plans as needed. Another cool invention that we have come out with is Kiro Powers. And we didn't go too much deep into it, but we did touch a little bit on MCPs in this conversation. Kiro is our agentic AI IDE that allows for spec-driven development. So what this means is if you download our Kiro product and you use it as your IDE for coding, and if you say, "Hey, let me build an agentic AI application that does a specific task," it creates the specs for you. It allows you to view the specs and say, "Okay, this is what I want to build." And boom, it gives you a detailed plan with these specs and it starts building it for you. It's really amazing, and I honestly suggest everyone that's watching this to go try Kiro. It's like mind-blowing, really. And what Kiro Powers is, is it provides you a native way to connect to MCPs. What a traditional IDE does is when you connect to all the MCP servers or any tool, it automatically loads it up. And then what that does is it takes up your cache memory and your RAM and the context windows. And it slows down the system when you're automatically loading all the tools. So Kiro Power, it caches in the MCP for you, and then it only invokes it upon use. So when you're actually using the tool. So it goes much faster. And it has native integration hooks with our other integration partners such as Datadog, for example, Dynatrace, just giving you observability provider examples. And you can also have Postman. So we have these cool integrations available and MCP connections available that you can call upon demand with Kiro Power. So I would say those two things are what we're doing in the future looking. And along with these inventions, I also just want to touch about how we are accelerating productivity and we're building these out-of-box services now that are based off of agentic AI to help our customers innovate to the next level so they don't have to think about building these things themselves. And I think that's where the future's going really is how can we build things more out-of-box for customers so they don't have to worry about, "Hey, what do I build with agentic AI?" They can just take this and run along with it.
0:32:38.4 Joseph Morais: I love it. Thank you so much for this conversation. A lot to be excited around the future of agentic AI on AWS, particularly with some of those new features that will help the developers and DevOps folks like I used to be and hopefully keep them from their PagerDuty going off. I can still hear that in the middle of the night.
0:32:56.7 Joseph Morais: Varun, I want to thank you so much for joining me today and sharing a wealth of knowledge with the audience.
0:33:01.0 Varun Jasti: Yeah, of course. My pleasure.
[music]
0:33:06.9 Joseph Morais: That's it for this episode of Life Is But a Stream: Cloud Connect. Thanks again to Varun for joining us, and thanks to you for tuning in. If you're on the early part of your journey of agentic AI and you're just trying to figure out how to get started, how do I scale, how do I observe it, how do I secure these things, and how do I get my agents the right data, a great place to start is with Confluent Cloud on AWS and looking at services like Bedrock AgentCore. If you'd like to connect, find me on LinkedIn. Tell a friend or coworker about us and subscribe to the show so you never miss an episode. We'll see you next time.
[music]