Did you ever walk into a conference session thinking you were ready for the week, only to realise the announcements were coming so fast that you almost needed an agent of your own to keep up? That was the mood across Las Vegas, and it was the backdrop for my conversation with Madhu Parthasarathy, the general manager for Agent Core at AWS.
He has spent the week at the centre of AWS's wave of agentic AI news, working on the ideas that are already moving from keynotes and demos into the hands of real enterprise teams. Sitting down with him offered a rare moment of clarity among the noise, and his calm take on what actually matters helped bring the bigger picture into focus.
Madhu talked through the thinking behind Agent Core and why he believes 2026 will be the year enterprises finally begin shifting from prototypes to production-scale agents. He walked me through the two areas customers keep coming back to, trust and performance, and why the new policy framework and agent evaluations could remove long-standing barriers to deployment.
His examples were grounded in real behaviour he is seeing inside large companies, whether that is internal support workloads, developer productivity, meeting preparation, or customer-facing flows designed to reduce the friction between intent and outcome.
We also explored the deeper shift introduced by Nova Forge, including the idea of blending enterprise data with model checkpoints to create domain-specific agents that can work with greater accuracy and context. Madhu explained why there will never be a one-size-fits-all model and how choice remains central to AWS's agentic AI approach.
My guest also reflected on how infrastructure changes, such as Trainium three ultra servers and expanded Nova model families, are shaping the pace at which companies can experiment, evaluate, and adopt emerging capabilities.
Trust surfaced again and again in our conversation. Madhu was clear that non-deterministic systems also introduce concerns, which is why action boundaries and guardrails are becoming as important as model quality. He described the excitement he is seeing from customers who now feel they have workable ways to give agents responsibility without handing over the keys entirely.
As he put it, this is the moment where confidence begins to grow because the guardrails finally meet the expectations of enterprise leaders.
We closed with the topic many people have been whispering about all week, modernization. Madhu reflected on AWS Transform, the push to help organisations move away from legacy architectures far faster than before, and the impact that agentic systems will have as they support full stack migrations across Windows environments and custom languages.
Madhu cuts through the noise with a grounded view of reliable autonomy, multi-agent orchestration, policy-driven safety, and the shift toward agents as true collaborators.
The question now is where you see the biggest opportunity. How might these agent-based systems change your workflows, and what would it take for you to trust them with the tasks you never seem to have time for? I would love to hear your thoughts.
Useful Links
[00:00:04] Have you ever had one of those moments at a conference where the announcements come so thick and fast, you wonder how anyone's going to be able to keep up? That was exactly how I felt having headed into my chat with today's guest, who has been right in the middle of AWS's wave of agentic AI news here in Las Vegas this week. And he's someone that's got a front row seat to how these ideas move from stage demos into real enterprise systems.
[00:00:34] But I wanted to hear how he makes sense of that shift. And after sitting down with me, gave me a chance to slow the pace a little and just understand the thinking behind it all. So as you listen today, I hope you find some clarity in the same places that I did. And hopefully if you've got time, let me know which part of this new software development is catching your attention right now.
[00:00:59] But enough from me. I'm going to officially introduce you to my guest live on the show floor at AWS reInvent right now. So thank you for joining me here at AWS reInvent in Vegas. Can you tell everyone listening a little about who you are and what you do? Yeah, I'm very excited to be here. And thank you for for for having me. I'm Madhu and I'm the general manager for AgentCore.
[00:01:27] I'm a long term Amazonian, have been in Amazon for over 15 years. And my background is building large scale distributed systems. And more recently, I've had a few gigs in a few AI startups before I landed back in Amazon to lead AgentCore. And I'm very, very excited to be talking about the future of AgentCore with you and your listeners. Yeah, I'm looking forward to talking about that. It sounds like we need to get you back on a later date to talk about your AI startups.
[00:01:55] That's a podcast episode for another day. But obviously this week, we've seen a huge wave of AgentCore announcements on stage. So when you look at everything unveiled so far, we will get into the details of some of those in a moment. But what is the shift that you think will have the biggest impact on how enterprise actually build and ship software next year? Do you see any big changes there?
[00:02:17] Yeah. So one one of the key reasons we built AgentCore was to help customers fill the chasm and the challenges from going from prototyping an agent to actually deploying them in production.
[00:02:35] And as we are working on some of the enhancements, and we'll talk about some of the recent announcements we made around making agents more trustworthy to be deployed in production, around making agents perform better and with higher quality in production,
[00:02:52] We are seeing a shift in customers being more eager to deploy their agents in production because of some of the undifferentiated heavy lifting work that we've done on their behalf. And I feel 2026 is going to be a year where we will see a lot more agents being deployed in production and operating at scale.
[00:03:20] So that shift is starting to happen right now. And here at AWS, they've been talking a lot around autonomous agents, how they can work for hours or days with very little intervention. So on this side of things next year, what breakthroughs do you think make this level of reliability possible? And what kind of workloads are you expecting customers to hand over first? Are you seeing any trends there in the kind of work that's being given to these agents, first of all?
[00:03:48] Yeah, so there are a few different kinds of workloads that we are seeing being more prevalent. One of the key ones is internal workloads that customers have where they want agents to reduce the amount of back office work that is being done by humans,
[00:04:12] enabling, freeing up humans to focus more on each business's area of focus. So, you know, customers are still figuring out ways in which they can leverage the agents. And an easy place for them to start is with some of these internal support-related agents, chatbots that can basically answer questions,
[00:04:40] meeting agents that enable customers prepare for their meetings, and so on. So that is one area of, one set of workloads that we are seeing customers look at. Another set of workloads is around enabling customers to, of, you know, customers' customers,
[00:05:04] so like users of the customers, to be able to more quickly achieve the outcome that they intend to achieve. No one browses the web to browse the web for the most part. They are there to perform a specific action. And the faster they can perform the actions, the sooner they can move on to, you know, to other parts of their lives.
[00:05:29] And that is where we are seeing a shift in customers wanting to make that easier for their customers so that the amount of time spent in achieving their outcome reduces. That one is one. And then they develop a productivity aspect of four existing enterprise customers. Like, as I mentioned, is another key area that customers are focusing on. And Nova Forge, that introduces the idea of opening training,
[00:05:55] giving customers access to model checkpoints, and the ability to blend their own data with Amazon Nova datasets. So how do you see that changing the balance between frontier model development and more domain-specific tuning? I don't see, with models, there's a one-size-fits-all. I believe that you're going to have, you know, and this is one of the things Amazon has always focused on,
[00:06:23] is offering choice to customers so they can choose the right models for their choice. And Bedrock and Agent Core and all of the other services that we've launched actually offer that kind of choice. And we believe that there exists a wide variety of use cases for domain-specific agents that, you know,
[00:06:47] enterprise customers would love to kind of have that would enable them to move faster and accelerate their agentic AI journey because their agents get that much more productive with domain-specific information. As I had mentioned, like before, one of the typical use cases we see is internal workloads
[00:07:11] for improving developer productivity and improving overall efficiency for enterprise companies as one of the use cases that agents are being deployed to. And there we believe domain-specific agents, domain-specific models will be valuable for customers. And that is the premise behind supporting and enhancing NOVA, announcing NOVA Forge for customers.
[00:07:39] We are very excited to see how customers are going to be able to use that. Another key insight here is that, as Matt announced in his keynote, your data is very powerful and it's everything. And with capabilities like NOVA Forge, now customers have the opportunity to take that data and apply it at scale
[00:08:07] to their agents, giving the agents that much more agency to be able to perform operations. And another theme across reInvent is trust. I think policy in agent call, agent evaluations and guardrails for action boundaries. Of all things, it surfaces ways to create a safer autonomy. So what does responsible agent behavior look like from AWS's perspective?
[00:08:35] The impression I get this week, it's something that you are taking very seriously. Yes. AWS has always been at the forefront of security and reliability and building trust with customers, be it their software, be it their infrastructure that they trust AWS with. And we take that very seriously. The agentic AI transformation is no different.
[00:09:02] We all know how fast the AI world has evolved in the last few years and how fast it is going to continue to evolve. Now, one of the key impediments for a lot of enterprise customers and just customers in general is the reluctance to see how agents behave in production
[00:09:29] because the power that agents have, their non-determinism that makes them powerful, is also cause for concern. And that is where the policy announcement becomes so much more valuable for customers.
[00:09:47] Because finally now, customers have a way in which they can control and provide boundaries around what the agents can do outside of the purview of the agents. Because that provides the level of determinism in terms of enforcing these policies.
[00:10:09] And that is inherent in all the capabilities that we have been working on and thinking about. And policies is no different. We believe there are two main areas of focus for customers, blockers or impediments for customers to deploy their agents in production. One of that is, you know, trust that the agent behaves as they expect it to behave in production,
[00:10:37] where it does not perform operations that could, you know, hurt the reputation of the company or harm them in any way. Policies provides strong guarantees around what is enforced. The other one is around how an agent performs. And that's where we announced evaluations today for ensuring that agents can actually behave and perform with high quality in production.
[00:11:08] In my many conversations that I'm having, many leaders tell me they feel stuck with legacy architectures that are difficult to modernize, the problem of technical debt, etc. And with AWS Transform, you're now applying agentic capabilities to full stack Windows environments and even custom languages. So how close are we to making modernization an almost background task instead of this multi-month project?
[00:11:35] And one of the things that I've seen here, obviously, with the you blew up an old legacy survey yesterday, which was incredibly fun to see, which the message I got was it's all about just finally removing this technical debt problem. One of the things we have seen time and time again in that we worked on improving for AWS customers is how can we help them innovate at a rapid pace that they are comfortable with?
[00:12:00] And that is where you've seen AWS focus on primitives or building blocks that enable customers to build. And one of the key impediments we've seen in the last few years is a lot of customers being hampered by their legacy systems and the time it takes to migrate them. And that is the premise of AWS Transform.
[00:12:23] And they're very heartened to see the order of magnitude difference in productivity increases. Like, you know, you heard Matt talk about like several thousands of hours and, you know, man, like person years reduced into like weeks of time to be able to let migrate legacy systems.
[00:12:46] And we will continue to push the boundaries of that with customers and hear feedback from them on things that we can continue to improve. And AWS Transform, you know, is one of the key areas of focus that will remain for us in helping customers move away from their legacy systems.
[00:13:12] And the conversation around agents often focuses on productivity. Yet some of the most interesting demos this year showed agents coordinating multiple systems and reasoning their way through multi-step tasks. Are there any new design patterns that you're seeing emerging as customers experiment?
[00:13:31] One area where we are seeing a lot of interest is in orchestration, like much more complex orchestration, given kind of how powerful agents have become to be able to support much more complex tasks where, you know, you leverage existing agents that were built to be able to interact with your agents and create the orchestration workflow that works seamlessly.
[00:14:01] One of our key focus areas with strands, which is a model-based framework, has been that we would want to support orchestration through model reasoning. And that is strands is different from other agentic frameworks in that regard, rather than have customers having to hard code their orchestration workflow.
[00:14:24] So we believe we can leverage the power of models to be able to perform, to reason about the orchestration steps, and then like build capabilities in strands, which we have built a few capabilities around swarming and graphs that enable customers to be able to orchestrate these multi-agent workflows seamlessly.
[00:14:50] And we are seeing a lot of interest in those kinds of orchestration flows where customers want us to reduce the barrier to entry and simplify the whole orchestration flow. And I think one of the most compelling things we've seen this week is the idea of agents acting as almost teammates rather than just tools.
[00:15:16] So when you look a few years ahead, how do you imagine that day-to-day work changing once agents can remember, reason, and collaborate across domains? I'm going to ask you to look into a virtual crystal ball. How do you see all this evolving in the future? Yeah, I have never gotten any kind of crystal ball answer right if I were to kind of reflect on that. One thing I do know for certain is that agents are here to stay.
[00:15:43] And agents, we will see as we remove these barriers to deploying agents in production, we will see a lot more agents be deployed in production and continuing to work effectively with humans on taking on some of the cognitive load of what humans can do,
[00:16:10] and also freeing them up to focus on their main task at hand. You saw Matt's keynote where a bunch of developers were able to leverage the capabilities with Kero's frontier agents to be able to solve some of the common problems and challenges that developers face.
[00:16:30] We will continue to see a lot of those agents continuing to evolve to become much more effective at what they're doing. And as we in AgentCore start to improve the quality of the agents and improve the trust and the security boundaries under which agents can operate,
[00:16:56] I expect to see a positive feedback loop that enables agents to become more autonomous, more safer, more secure, and more reliable working with customers. Working for or on behalf of customers. And also at the event this week with Tranium 3 Ultra Servers and the expanded Nova model family,
[00:17:22] the infrastructure side of things is undoubtedly evolving at high speed too. So how tightly do you think the future of AgentC systems is tied to the hardware progress? And what does that mean for organizations that just want to scale fast? I believe that the evolution we're seeing with hardware and model evolution and innovation is going to continue.
[00:17:46] And it is going to get better for customers to be able to become more efficient, more cost effective, and more secure as well as we start adding more and more capabilities in the infrastructure layer for customers. Now, one of the areas that we believe we can help customers in this journey,
[00:18:16] because this evolution is happening and it will continue to happen, is to enable them to move, to experiment with newer capabilities that come in and choose the options that work best for them and migrate to them. And AgentC course evaluation is a great example of how customers can do that.
[00:18:45] You can choose a newer model and have your agent use the new model, compare it with kind of the existing model performance, and be able to evaluate within hours how the newer models perform for your use cases, and then easily move. So today, without evaluations, that process takes you several months to kind of,
[00:19:14] ooh, by which time a new model version has emerged. And we want to accelerate that for customers without them having to do a lot of heavy lifting. And that is one area where we will continue to enable customers to be able to keep up with the evolution of hardware and infrastructure and all of that. And finally, obviously this week, you've been meeting with customers all week
[00:19:42] and reporters here in Vegas and attendees on the show floor. What is the most common concern or misconception that you hear about Agentic AI? And is there any reassurance or guidance you think leaders need most as they begin adopting these systems? Because any of us that scroll down our news feeds, we see the good things, the bad things, and the bad often seems more than the good. But I think that's not true at all. But for the conversations that you're having, what's the general consensus here? Yes.
[00:20:11] So the main thing I'm hearing from customers that is very heartening to us is the fact that they are able to understand how with the launch of policies for agents, they have a way to actually deploy these agents in production, especially the aspect of policies that provides that determinism
[00:20:40] for a non-deterministic system fundamentally. That shift in thinking is apparent when we talk to customers as they're hearing about the policies announcement that came up. And to us, that is a good indicator of customers wanting to try out these new capabilities that we want to try and be confident that they can deploy them in production.
[00:21:11] And we are also excited to see how customers will leverage policies in ways that apply to their use cases and see how they can continue to move the shift from prototyping of agents to productionizing agents and how we can actually help them in that journey. I think that is a great moment to end on today.
[00:21:38] I will include a link to your LinkedIn for anyone that wants to carry on the conversation with you. But other than that, is there any way you want to point anyone to who's just interested in keeping up to speed with some of the announcements coming out this week and where you're heading in the future? I am personally incredibly excited about the journey that we are taking in the agentic AI world and enabling customers to provide them with the best place on earth
[00:22:06] to build, deploy, and manage agents. I would encourage all your listeners to, if they have five minutes, try out AgentCore. That's all it takes. We've done a lot of work in improving the developer experience and also leverage Kiro to build an agent and deploy it in AgentCore. We have an AgentCore MCP server that they can point Kiro to
[00:22:35] and simply use prompts to build an agent and deploy an agent to AgentCore. I'm very excited to see how customers are going to leverage this to build agents that will enable them to be more effective and productive for their use cases. I think that's such good advice. We all read about so many things and get excited about it, but to anybody that's interested in any new technology, I always say, have a play with it. Get hands on with it. Absolutely. That's the best way to learn about the tools
[00:23:05] and see how it works for your use cases and what works and what doesn't. Awesome. Well, once again, thank you for stopping by and sharing your story and all the great work you're doing. Thanks again. Thank you very much. It's been a pleasure. I think listening to my guest today, breaking down the thinking behind AgentCore left me with a much clearer view of where this path might lead. And yet, the pace of change at reInvent can feel overwhelming.
[00:23:31] Yet, he was able to bring a steady voice to topics that are shaping the next generation of enterprise software. And his comments on reliability, modernization, security, trust, these are what things customers expect from these systems as standard. But I'm curious, what ideas do you think will matter most as agent-based systems move into our daily workflows? And where do you see the biggest opportunity?
[00:24:00] And what questions does this raise for your own work, for your own organization? As always, let me know. TechBlogWriterAtOutlook.com LinkedInX Instagram Just at Neil C. Hughes And remember, drop by TechTalksNetwork.com where you'll find all of my podcasts and over 3,500 interviews. But we're out of time for today. I'm going to get back on that show floor now and see if I can find somebody else to speak with.
[00:24:29] But I'll be back in your podcast feed, same time, same place tomorrow. Speak with you then. Bye for now. Bye for now. Bye for now.