The State of AI Agents in 2025 & How to Use Them

Welcome to the Next Wave podcast. I'm Matt Wolf. Today I'm joined by Deepak Singh, the vice president at AWS, where he leads the teams building Amazon's Agentic AI infrastructure. Deepak has been with Amazon for over 17 years and has been working with neural networks since long before most people.

Had even heard of AI.

He's played a key role in shaping how developers use generative AI at scale and and now he's focused on building the next generation of AI agents. So in this episode we dig into what actually makes something a true AI agent versus just another buzzword. How AI agents are already transforming industries from Formula one racing to drug discovery, and where this technology is all headed next, from multi agent systems to new standards like MCP and A2A. We'll also explore how businesses of any size can start using agents today, and what it all means for job developers and the future of work. If you want a clear insider's look at how AI agents are going to reshape business and technology, you're going to love this conversation. So let's just jump right in.

Using only 20% of your business data is like dating someone who only texts emojis. First of all, that's annoying, and second, you're missing a lot of context. But that's how most businesses operate today, using only 20% of their data. Unless you have HubSpot, where all the emails, call logs and chat messages turn into insights to grow your business. Because all that data makes all the difference. Learn more@HubSpot.com.

Hey Deepak, thanks so much.

For joining us on the show today. I'm really excited to chat with you about AI agents and Amazon's role in this whole agent world and all of that kind of stuff. So how are you doing today?

I'm doing great. There's a lot of interesting stuff happening at work and in the space of AI agents. So excited to have this conversation. Awesome.

Well, let's get into a little bit of backstory.

I don't want to go too deep on the backstory because I want to get into like the meat of talking about agents, but I would love to learn a little bit more about how you got involved in Amazon, how you got involved in AI and the agentic world. What's the sort of backstory there?

Yeah, so I've been at Amazon aws for over 17 years. It's still the only tech company I've ever worked at. Prior to that, my life was all in mostly biotech and scientific software. I have a PhD in theoretical chemistry. It's quite interesting. My first job after my PhD was actually writing AI algorithms for predicting protein structures. Many of those techniques are completely ancient by modern standards, but the first time I wrote neural network was over 25 years ago. Over the years, I've become more and more interested in two things. One is developer experience, and the other one is how can we use the power of computing to make the lives of developers much easier. So as generative AI became popular and became a thing in the last three years, it was a natural move to start trying to see how generative AI could impact how developers work day to day. So. So that sort of combination of that interest is what brought me into the Genai world and agents world. So I like to say my team is a team that cares mostly about developers and developer experience. And generative AI is the most powerful tool we've ever had to radically change that. So that's why I'm doing what I do.

It's so interesting because most people have seen how AI has exploded over the last, you know, call it three years, maybe 2021, 2022 is when it really, really started to bubble up. But really, I feel like before that nobody really had much interest or was talking about AI, but obviously there was a lot of progress behind the scenes, a lot of things happening. From your perspective, what do you think it was like the biggest, like, pivotal moment, that shift that got the world on board?

Yeah. So machine learning, I'll call it machine learning, has been around for a very long time. You know, along the way, people have called it AI, and depending on how you think about it, people have said, oh, it's oversold or doesn't work. But machine learning has always worked. It worked 30 years ago. The ability to learn and then predict has been there from as long as I've been working with computers. And those techniques got more and more powerful. Several years ago, there was a lot of excitement about deep learning. Part of the reason we don't talk about deep learning is it's there everywhere. And it's almost accepted that it's there everywhere in everything that we do, whether it's how we shop or with how we use the Internet. But I think what made generative AI such a big deal and why people saw it and were able to get excited by it in a much more general way is because you did not have to have a PhD in computer science to really appreciate it, because you could go and type a query into a chatbot and immediately start getting answers. We are at this point in life now where when I'm planning a vacation, I generally tend to use some AI tool to figure out where I want to go and I want to set up my itinerary. So the fact that it has an impact on your day to day life is what I think made it so impactful. And I think that started off in the last three years with the great customer files called Perplexity, with the kind of search engines they're building or you know, back in the day with ChatGPT and the consumer side of it. And of course the fun part is that one of the first use cases for generative AI was outside of the pure consumer space was developer tools was programming. Part of me thinks it's because the people building these tools were programmers themselves. But it's also a problem that's particularly well suited for generative AI. So you know, you combine the excitement that the world has with these tools and the fact that developer tooling was the first area of interest makes it even more exciting for folks like me.

Yeah, no, totally. I feel like the world really started to take notice around ChatGPT. Right end of 2021, ChatGPT came out. That was sort of a very easy consumer entry point point into AI and I feel like that was kind of when everything really kicked into high gear. You mentioned machine learning and you mentioned deep learning and I know it's a very sort of nuanced, complex topic, but can you sort of give me like the layman definition of each? Like what's the difference between machine learning and deep learning?

Yeah, I'll keep it very simple. I actually differentiate ML with generative AI more than anything else and sort of what differentiates between the two. Historically when we've talked about machine learning, we've talked about predictive modeling where your system learns and it's able to predict based on statistical analysis. Like these are statistical systems and they're using neural networks and other techniques to predict what the next step might be. The most probably well known examples of these are things like when you start typing an email and the next few lines get written up, or in the area of computer vision, your self driving cars and things like that all use those kinds of techniques. They're using deep learning techniques to figure out how to react in a particular condition. They're making a statistical analysis of the situation. The great part about generative AI and why people get so excited about is it is generating information not based on some common math, but they're probabilistic in nature. They try and be for the lack of A better word, they're creative, which the traditional ML techniques were not. They were strictly mathematical and statistical. Generative AI allows you to create. My favorite example is write a poem with these topics and write it like Wordsworth, Shakespeare and whoever, those kind of fun apps that you show your kids. So the ability to generate from all the learning and training that has happened, but try and come up with something novel. And that's what made generative AI so exciting. And of course we found other ways of applying it, but that's the big difference between the two, at least from a layperson's standpoint.

Yeah, and it totally makes sense because, you know, we also get what the industry calls hallucinations. Right. Which kind of comes from that thing of trying to predict. Right. It doesn't necessarily always know the answer, which could be a feature or a bug. Right. If it doesn't know the answer and it just makes up an answer, we call that a hallucination. If it doesn't know the answer, but it comes up with something novel that nobody else has thought of before, put something, you call it creative now it's creative, right?

Yeah, exactly.

But to me that's been really, really fascinating to watch.

But.

So when it comes to deep learning versus machine learning, my kind of understanding, and I don't know, very deep, I don't have like a computer science degree or anything, but my understanding is all deep learning is machine learning, but not all machine learning is deep learning.

That is correct. ML is a very broad space. You can just use simple statistical approaches to things like DNNs and CNNs and advanced neural networks, which is all a lot of what deep learning is.

Right, right. And I don't know if this is like an accurate description, but the way I've sort of heard it is like deep learning is when you sort of give an input and it passes through a whole bunch of different neural networks before you get the output. That's what makes it deep is the fact that there's a depth to the amount of neural networks that it passes through.

Correct. I mean, there's a lot of unlabeled data which. And you go through layers and layers and in many ways what the biggest difference between the neural networks that, like the ones I worked on 25 years ago, there was like a single layer of feed forward neural networks because your compute power was limited. And now your compute power is quite significant. And that allowed you to do deep learning over GPUs and start building these multi layered systems. And that's actually what's allowed you to train the large language models. I mean, language models have been around for a long time as well, right? It's large language models, these frontier models that are new, that are able to then give you these probabilistic responses that make them feel so magical.

Totally makes sense.

Everyone everywhere is talking about AI agents right now. But here's the thing. Most companies are going about it all wrong. This guide cuts through the hype and shows you what's actually working right now. HubSpot has gathered insights from top industry leaders who are implementing AI agents the right way. You'll discover which agent setups actually deliver ROI and how businesses are automating their marketing, sales and operations without replacing their teams. Get it right now by scanning the code or clicking the link in the description.

Now let's get back to the show. Well, let's shift gears into agents a little bit. And when it comes to, like AI agents, I feel like the definition has been kind of fuzzy.

Right?

Like, I feel like you can ask 10 different people in the AI space and their definition of an agent might be slightly different than other people's. How do you or aws, how do you guys sort of define agents?

Yeah, an agent. In the end, we see it as a combination of a set of things. One or more frontier models underneath the hood. You have to have an LLM. That's the heart of an agent. An agent is connected to tools, a set of tools that can use and that's actually been one of the biggest advances in agentic systems is the ability for these LLMs to use tools to data, often like a RAG system or through MCP to other things. So an agent is a system with an LLM at its heart that then you interact with to do anything from write software, to create images, to check your email, to do customer service. But the key part is an agent is given an outcome. It has access to tools and given a goal and the tools and data it has access to, it's able to autonomously figure out what it needs to do to accomplish its task. A great example of, I think the current state of the art or in some ways of Agentix systems is an example that we have from somebody inside Amazon where they posted a slack of they were working with about 100 files. They needed to make some changes to all those files and they gave it a prompt, make this change to here's all the files. The agent started making those changes. They made changes to three files. It recognized that had done the same thing to three files. And it had 97 other files to make that change to it wrote itself a script to go and do the other 97 in an automated way. That ability to autonomously make that decision is what makes agents interesting. The way we build them now, it'll actually ask the user, hey, I want to do this. Are you okay with it? And that autonomy is what makes an agent an agent.

Gotcha. Yeah. Because there's a lot of tools out there, right? You've got like, you know, make.com and Zapier. Zapier. These tools that can use AI within a workflow, but they won't necessarily realize if they made a mistake and then sort of go back and kind of double check themselves. Yeah, but I feel like maybe when we're talking about like a real AI agent, a real AI agent going to do the work, double check, did I do this right? Are there any issues? And if it finds anything, sort of double back and try again.

Correct. Yeah. We call that reflection. It's able to reflect on its own work, go back and then maybe correct it. If it sees a bug, it then goes back and tries to figure out a different path to solve that bug. It's not uncommon in software development, for example, to just post your error log into the agent and then the agent goes, okay, I think I know how to fix it. And it'll keep trying until it does, or if it can't, it'll ask for help. And again, this level of autonomy is what makes agents agents. I think part of the challenge out there and why it gets so confusing is people will label a lot of things agents, but the autonomy part, the ability to track, go towards a goal for us, those are fundamental capabilities of an agent.

Right, Right. Yeah. I feel like agent's kind of a buzzword. So you get a lot of software companies out there wanting to toss agent in as a selling point when maybe it doesn't totally meet the sort of true definition of an agent. So how have you seen agents evolve over the last few years? In my mind, an agent, ideally we get to a point where you give it a task and you can just walk away from your computer for a couple hours, come back, and the task has been completed. I feel like that's where we want to kind of get to, but most agents I feel like, aren't there yet. Like, but how have you seen that evolution of agents evolve to where we are now?

Yeah, I'll actually go even one step further and go back to the sort of early days of LLMs where there were single shot, you sent a query and you got A response. Now, somewhat, very quickly you started getting these multi turn agents, we'll call them, the early agents were basically multi turn conversations. These chatbots that you started seeing appear everywhere. You gave it a query, it gave back a response. You might say, yeah, but I wanted you to think about it this way, not that way. And they came back with a response and they had memory, they kept context of what was going on. But LLMs got more powerful, they got better reasoning capabilities, they got the ability to use tools. And I think starting somewhere in the middle of last year, at least in my mind, realistically, Agentic AI started about a year ago.

And what made Agentic AI possible was LLMs, which you could just drive. And they had this ability to go use tools to reflect, look at different parts. And you had to give them an outcome, give them the right context. The context could come from, here's the data you should care about, here are some of the APIs and tools you can use. And they were able to then make decisions based on that. I gave you one example already, but I think that's what's allowed agentic AI to evolve from just chatbots to meaningful systems where they're able to do work. They could be doing software development, they could be agents that help a business analyst get work done where they can, you know, it can create a dashboard or something like that. You know, a great example is Formula one used to have this problem where, you know, they're collecting data from their machines, from all their cars, right? And let's say there's a technical issue. They had to create a root cause analysis, like to figure out what was happening, because I don't know how familiar you are with Formula one, but one of the beauties of Formula one is between one race and the other. The car changes because they look into it and they change their aerodynamics, they change their engines, et cetera. So if you have an issue in the past, it used to often take three weeks to analyze all the data and figure out what to do. Three weeks is a long time. That's three races in Formula one time. So they created this root cause analysis assistant using Amazon Bedrock agents and that was able to reduce resolution time significantly, about 86%. So their triage got cut down and overall they were able to resolve the issue in three days. Now you can make changes for the next race, which in Formula one could be the difference between driver's championship or not, right? So that's an example of the kind of things you can have an agent do. And it's that level of impact, I think that makes agents so exciting to all of us.

Yeah, absolutely. Now you mentioned that you feel like agents kind of bubbled up onto the scene about a year ago. Is there like a moment or a tool or something that happened that you feel like was sort of that initial catalyst?

Yeah, I would say the key catalyst is the rise of what you will call thinking reasoning LLMs. Because prior to last year LLMs couldn't reason and then you got the generational reasoning LLMs and those made a huge impact. So I think that at least to me that's like the moment in time that took agents from just being multi turn assistants to being truly agentic AI that was able to do work autonomously. Some of that you could do before, but that just took it to another level.

Yeah, yeah. I mean you mentioned perplexity a moment ago and perplexity is one of the first times that I feel like I started to get like a little bit of a taste of agentic AI because you would ask it a question and then you would see it actually sort of make a checklist and then work through and like check off the things on its checklist before actually giving you the response. And for me that was when the concept started to click of like, okay, when you're using a chat GPT or cloud or something like that, right. You give it a prompt, you get your output back and you just kind of go back and forth. But now when I first saw Perplexity doing this, it was like you give it a prompt and it sort of plans out a set of actions and then it follows through on all those actions. It may not have that. What did you call it, where it sort of recursively checks back. Yeah, it didn't have that yet, but. But you kind of started to get that glimpse of like, okay, it's actually going through and doing a whole bunch of tasks for me instead of just the one task of responding to my prompt.

Yeah. So I'll actually use the example of kiro, which is an agentic integrated development environment that we released in July, and KIRO as an agentic ide, that's how it works.

Right.

So typically if you go back even to the early days of developer assistance using generative AI, what they did was auto completion. You started typing and they would try and finish your function or whatever. Then you got to the next generation of assistants. It's a more chat based where you're able to say, here's my code base, I want to write a function that does X. It will look at your code base and say, here's the code for that function. Now, with agents, you can actually drive that. They pretty much write all your code. But here's where what we decided to do was a UX that made it much more powerful. And I think it gives you a good idea of how agents behave. So using Kiro as a developer, I might say, hey, I want to build an app that does X. I want an app that build a shopping cart or visualize proteins or something like that. Pick your problem and you are talking in natural language. This agent collaborates with you to create a set of user requirements and a design document and dependencies. Then what it does is it creates a list of tasks that it needs to do to complete its work. It's going to say, I'm going to use these libraries, I'm going to use these tools. This is how I'm going to do it. As a user, you can just keep hitting yes, yes, yes, or you can say, now I want to make some changes. I think there's a better way. You use your judgment, then you hit go and then it goes ahead and does it. But with this chain of thought reasoning that these modern LLMs have and modern agents have, you can see it getting stuck. For example, in Kira, the first thing the agent does is writes a test for itself. Because you want to write a test, you're doing test driven development. And if it fails the test, it goes back and tries to solve it because it learned something from the error that is, I think, true agentic behavior. And I think over the last few months we have started to see people doing such meaningful work with it and I think it's only going to evolve from here.

Yeah, absolutely. Looking forward a little bit, you know, mentioning evolving AI, where do you think it's all headed? Like, where do you think agents are going to be a year from now? Maybe five years. I have a hard time seeing five years for the future. I have a hard time actually guessing what we're going to see next year. I'm usually pretty wrong. I'll say this is coming next year. And then the thing I predicted is coming next year comes three months later. But I'm just curious if you have a vision of where you see agents like a year from now or maybe five years from now. Where is this all headed?

Yeah, five years is really hard in this world which evolves so quickly. I have this joke that anytime I present a document, it's obsolete by the time I get back to my desk because of how quickly the world of AI moves these days. I think there's two or three areas that it's going to move. One is even the agentic systems we have right now are somewhat limited. They tend to be a single agent, not multi agent applications. At most they'll have two or three agents in them. You're going to have applications that are connections of agents, like lots of agents talking to each other, asking each other for help. In an autonomous world as well. As we drive more and more autonomy into the systems, you'll get these more complex applications. And your tooling to do those needs to evolve as well. A couple of months ago, we launched something called Amazon Bedrock. Agent Core and Agent Core is a set of primitives. But the idea is somewhat simple, at least conceptually, which is you need a core set of primitives to run agents at scale. You need secure, safe runtimes where agents can be isolated from each other and run in isolation. You need a way to identify an agent. An agent isn't a human, but it's not a traditional service either. It has a difference. How do you give an agent an identity? How does that agent communicate with the rest of the world? And by the way, that may not even be a digital world. It may be running on a robot where it needs to communicate with the physical world. And then how do you go and audit what's happening to these agents? So, and the protocols and MCP and A2A are all improving underneath it. This is the boring part of running agents. The operational infrastructure and capabilities of running agentic applications at scale is going to explode in terms of maturity over the next year, right? Yeah. Agent Core is two, three months old. And that shows you it's just getting going. But by next year it'll be very mature, it'll be generally available, and people will be running large scale agentic applications on something like Agent Core. That's one.

What do you think the bottlenecks are right now to, you know, being there?

I think the big ones are just understanding. Six months ago, there's 20 ways of doing these things. Now there's a set of sort of standard protocols that are developing so at least two agents know how to talk to each other instead of you having to write bespoke code. Most of us are handcrafting runtimes to run these agents. Now they're standard ways of running these runtimes. So the barrier to entry gets lower. Right. You don't have to be a team that's an expert at building all of these components. You can use managed services from somebody like AWS to go run them. So the barrier to entry for a customer or any business that wants to build these agents goes down significantly. So that's one. The second one, and I think this is the one that most people talk about when they talk about agents, is the level of autonomy. Right now for the most part, you still have humans in the loop. A lot like in software development, you have people driving the behavior of the agent, you're prompting the agent and even if it goes and does tasks in the background, you still have a human. They don't run for two days, right? They'll run for 10 minutes, 15 minutes, 20 minutes, maybe a couple of hours. But you're going to have systems that run much more autonomously. When you're doing that, a few things happen. One, you have built trust and the quality has to be there. So you need systems that can verify and validate along the way. So that becomes even more important because you don't have eyes on the problem. So at aws, we believe in neurosymbolic AI techniques which help you verify the behavior of an AI system, make sure it's giving you the results that you want to, that are safe and so on. There's many techniques, that's one area that we are heavily invested in. And then how does this all interact with each other? So the level of autonomy, I hate to say this kind of almost like self driving car world, right. In some ways the level of autonomy is going to go up and when that happens, how businesses, companies are integrating AI into their day to day work changes. So I think that's going to be the other big one. Beyond that, your I guess is as good as mine.

And what do you see as Amazon and AWS's role in all of this? Obviously you guys have bedrock, you know, you've got your cloud services and things like that. What is AWS doing to sort of progress agents forward?

Yeah, and I'll go back to agent core and bedrock. In the end, to run an agentic application you need great agentic infrastructure. So at the most core level, we want to give you the best agent infrastructure to run your agents. That includes something like bedrock that gives you LLM choice so you can pick your LLMs. Then you need a place to run your agents, which is agent core runtime. So this is, it's serverless, it's secure, it's isolated. We've made a lot of innovation on the micro VM side. Like these are innovations that we've done over the years, like the nitro System and micro vms. They're taking advantage of them to run agents in these isolated sandboxes and scale them. We support any LLM, any agentic framework. You can write your agent however you want it. We want you to run them at scale. We are investing in Agent Core Identity. So how do you identify an agent? What identity does that have in Agent Core Gateway, which allows you to connect agents to the rest of the world. Agent Core Memory. Agents have memory short term and long term. They have to remember what they did to reflect and make decisions. And at the end of it, you want auditability. Well, you want observability for two reasons. You want to respond in real time. If you have access to all the metrics and all the logs, you can do that. But later you can also go and see what are all the decisions has this agent taken, what did it interact with, at what time did it do so? Because those have become important. So at a base level, we are building the most high performance, scalable and lowest cost infrastructure to run your agents. It's what AWS has been doing for applications for a very long time. The second area that we are heavily invested in is building agentic applications ourselves. So for me that will mean agents. To help you build software, we are increasingly using agents in other areas. For example, we have something called Connect for Customer Service, which increasingly has agentic AI capabilities inside it to help customer service agents and customer, you know, virtual call centers be more effective? And those are good examples. We are building these AI agents into the fabric of how AWS operates so that our customers can benefit from it. They don't need to know anything about agents, they're using them.

How do you see Amazon and AWS approaching responsible AI? Like, is there specific things that we feel like the human should always be in the loop on, or do you think we're going to get to a point where we're totally just trusting this to go and handle the task we give it?

Yeah, very specifically today, for example, if you do software deployment, we don't have any tools out there to help you do well. We have tools to help you, but we have no agents that will deploy software for you. We want the agent in the loop. There are other areas where we'll give you the tools and recommendations, but you as a company can decide whether you want to apply those or not. And I'm speaking at a fairly high level right now. For example, if you're writing software with an agent, our recommendation is, and we give you tools to do that, is to do code Reviews, but you may be a two person startup that goes yolo, I'm not going to do code reviews, I'm going to trust the agent and just merge this code in and ship it. But that's the decision you make. We'll give you our guidelines, but deploying software, we don't give you any agent that helps you do that. Today as an example, the area that we are really, really invested in is so Bedrock has Bedrock guardrails that if you are building an agent is actually something that a lot of folks don't always realize. Based on my conversations with them. You have to make sure that the prompt that the end user is submitting goes to a set of guardrails before it goes into the agent, right. So that you make sure that there's nothing malicious going to happen. And then on the other end you need to apply guardrails to make sure that the results that are coming out meet your company standards, don't have anything malicious, et cetera, or toxic in there. So those set of guardrails, we publish them and with Bedrock Guardrails we allow you to tune it so that you can do it based on your policies, et cetera. For example, we use for our own applications we will use LLMs that we believe are a high standard themselves that have built in guardrails, but then you can apply layers on top of it. That's one area. The other area is your favorite one. Hallucinations. How do you make sure, like I'm a company that needs to provide factually verifiable information and you as a consumer asking me for the price of something, how do I make sure that I give you the real price and not a made up price? Right. Hallucinated price. So we have a new feature as of two months ago in Bedrock called automated Reasoning checks. These automated reasoning checks use mathematical formal models that you can create to verify that something is actually correct. So for example, if you're asking me a pricing question, I'm the agent, we can build these checks into it which make sure that when I'm giving you a pricing response, it's a factually correct response. This goes back into these automated reasoning and neuro symbolic techniques that we invest in very heavily. So in addition to the guardrails around, how do you make sure your agent is not using bad language or using style that is not appropriate for your company? You can also do things along lines of making sure that the accuracy of the responses is very high with using some of these other techniques. So those are the areas that we Invest in ourselves quite heavily, both from a research perspective, but also product perspective.

Gotcha. Yeah. So basically there's guardrails. So when somebody enters the prompt, it kind of passes through the guardrail before letting the LLM even see what you prompted. And then when the LLM gives a response, there's another guardrail that that response passes through before the end user sees it. And you can kind of tailor what does or doesn't make it through.

Yeah, there's some that's built into the LLM. Some. A lot of LLMs are. For example, some of the products that from my team use Sonnet 4. Sonnet 4 has built in guardrails that are pretty good, but we put additional guardrails on both sides to make sure that the quality and responses are appropriate.

Yeah, makes sense. There's a couple terms that have been brought up a couple of times throughout this interview, and I'd love to help sort of define them for people listening that may not understand them. You mentioned MCP model context protocol, and you mentioned A2A, which is, I believe.

Agent to agent, agent to agent.

Can you help sort of define those and what they're used for within the agentic process?

Yeah. MCP is the world's best hammer. Everything looks like a nail. I'm only half joking. Even nine months ago, when sort of these reasoning LLMs came, one of the first things that we observed was that LLMs were now capable of using tools. They have been for about a year and a half. And the tool could be ability to read and write files, for example, about a year ago, maybe a little more, we published a blog post about something called TextEdit. And TextEdit was just a text editor, like an editor that we had made available to the LLM to read and write files, to look at text, to delete it, to change it. And that's the tool. So people are writing tools for their LLMs and their agents in a very bespoke way. We have an agent framework called Strands, which, before we released it publicly, had its own tool system, like how to make tools available to it. But then MCP came around and what MCP is just a standard protocol to make tools available to an LLM. And what's happened is it organically became very popular to the point that pretty much everybody supports it now. So if you want to give your LLM access to anything, whether it's an API, another agent, an SDK, a set of tools that you've written, you can wrap them in an MCP server and make them available to the LLM. And the great thing on MCP is those tools can be any tool. As long as somebody is making it available as an MCP server, it is now available to your LLM as long as you're giving it permission to access it. So MCP in the end, is a standard protocol for tool access. That's how to think about it.

Right? So you have all these software products out there, right? All sorts of SaaS products that you might be able to develop something and tie into it via. Like an API. Right. But the problem is everybody's API is sort of developed different. They all have their own documentation. There's no real standardized API. Everybody kind of builds their API differently. And so when you want an LLM to go and talk to some of these tools, telling it to go talk to the API, there's no sort of universal way of doing that. And I think mcp, my understanding solves that. Right.

It kind of creates this reverse.

It wraps it. Yeah, yeah, it wraps it in a way that an LLM can understand. So anyone can now write a tool. The reason I say it's a worldwide hammer is it's like if you talk to anybody about I want to do this, that's like, oh, just put it behind an MCP server. So it becomes sort of. It's almost a half joke in the industry now. But it's a very powerful tool. It's a really powerful hammer. And it's changed the way we think about how agents are written because it's a standard.

And then what about agent to agent? How is that different from mcps? You know, how does that fit into the workflow?

MCP is about tool access now. MCP is going beyond tool access. There's more capabilities coming into it, but that's where it started. Agent to agent is a protocol. It's like a wire protocol for one agent to talk to another agent. So how does agent A know how to talk to agent B? And if you have a standard protocol to do that, all agents can talk to each other. So I think of it more like a networking or sort of wire protocol standard. Like you got things like proto buffers and thrift back in the day that allowed two services to talk to each other. It's just a communication protocol between two agents. That's the best way to think about it. I don't know. Agent to agent communication is so early. I think that's how A2A and standards like that will evolve is less clear to me. MCP is pretty much ubiquitous at this point of time, right? Right.

So I want to talk a little.

Bit about like real world applications of all this stuff. So this show actually has a fairly large B2B audience. There's a lot of people listening that they want to know how do I leverage AI, how do I leverage agents, how do I leverage some of this technology into my business to make things easier? So there's a couple things that I want to sort of go down with you real quick. One of them being can you share some real world examples of agents you've seen? And then after that I would love to sort of dive into. If somebody wanted to go build their agent, can we sort of walk them through? Here's what you need to go do to do it. But first let's start with some like real world examples.

Yeah, so my favorite one, maybe because of my past life, is one from Genentech. Genentech worked with us to build set of agents to improve their drug research project. They used Amazon Bedrock Agents. Amazon Bedrock Agents is a great way to build simple agents. You know, you can't mix and match too much stuff. But it's gives you a very simple path to building an agent. And what they wanted to do was they have to all these therapeutic areas and they collect millions of biomarkers and they need to validate them. It's actually a very manual process, having been in the industry. So they built an agent that automates biomarker validation. You know, as scientists generate these biomarkers. Well actually they're coming off instruments before they decide which ones should be sent on for more experimentation or to build a new drug. They're to validate them. Is this even real? Does this even make sense? And so they built an agent along with us to improve this process. And they say automated about five years worth of manual work by doing that. And so this helps them do drug target identification, improves their research agency and so on and so forth. And a more sort of consumer application is DoorDash has AI agents in their contact centers. You know, so they do for voice and chat to answer frequently asked questions. I mean I don't use DoorDash that often, but when I do, if there's a question, I suspect there's hundreds of people or similar questions. They have agents that become very good at answering those. That key part about a gen agent is it's not a set of canned answers. They're responsive. As I said, they can react right, because they have access to my order, they have access to my data, they know what I've done before, they can give me an answer that's relevant to me. Right. So they're able to field hundreds of thousands of calls per day. They can actually respond to dashers. That's the main one. When a dasher has a question about what they need to do about an order or they need to clarify, they're able to get responses back in like two and a half seconds. They don't have to go look for somebody who might know the answer because they have access to their own system and the dashers, what the dasher is doing to the order and so on and so forth. So these are examples of the kinds of agents people are building. And there's obviously from my perspective, the best examples are the agents that we have built ourselves for software development which are allowing people to build powerful systems and change the productivity of engineering teams. Because you're using agents, right? Software.

Yeah, that makes a lot of sense. And I imagine with like the doordash example it's either fine tuned with all of the past sort of customer support queries so it's seen how all of those have been handled, or they're all, you know, dumped into a document and retrieved via rag or something like that.

Right.

Like they're able to actually see all of the past history, see how everything was responded to in the past. And there's probably enough data there that the AI could respond pretty dang well based on everything it's seen throughout history.

Yeah, and with modern sort of contemporary agents, it's not just the fact that they have access to the data, it's what they do with the data. So the data is part of the context. I often say context is king. The better your context, the better the agent does. But because as you said, they have good context, their agent's able to be very useful to the end user because it is able to interpret natural language questions so much better and take action.

Right, right. Yeah, that totally makes sense. Now when it comes to. Let's just pick a random example. Let's say there's an online E commerce business and they want to set up some sort of agent to help them out. What's their first step? Like what do they go do to set up their first agent?

So I'll give you my recommendation because it is kind of amazing how easy it is to use in some ways, especially for a developer. Agents need a framework to build the agent frameworks are, you know, lack of a better word. It's their libraries or some kind of development framework that makes writing an agent easy and allows you to ship it. There's been frameworks have been around for a couple of years. The early frameworks were quite complex because they had to work around the limitations of the LLMs at the time. They had to do a lot of the heavy lifting themselves. But as LLMs have become more powerful and significant, you had to build chain of thought outside the LLM. But now LLMs themselves are capable of reasoning. Right? So what becomes more important is how do you drive the behavior of the LLM take how do you make help it reason better? So we built something which we've since open source called the Strands SDK. So the library, I think it may be in more than one language. Now the original one was just in Python. The Python library that you use to essentially drive agentic behavior to what we call an agentic loop or model driven development. And you put into your code, you describe your agentic behavior and Strands has access to tools, it has access to MCP, it has access to A2A. It's able to help you build that application very, very quickly, just the way you would build any application as a software developer. If you want more the drag and drop, I want to just pull things out of it. Use something like Bedrock Agents. It's a drag and drop system. And then when it comes to how you want to run your agents, that's where Agent Core comes in. It feels just like any other AWS service. It has APIs, it has a console. You've built your agent in Strands, for example, you don't have to use Strands, but let's say you have. You can deploy it on Agent Core runtime. You can use Agent Core Memory for its reflection. If you want to do reflections, you can use Agent Core Identity to give it an identity you want to connect it to. 50 other things. Use Agent Core Gateway because that gives the connectivity and also the ability to manage traffic, all the things that you want to do. So from an AWS perspective, what we would say is use Bedrock to host your. You know, pick an LLM. You can use any framework, but Trance is just open source. You can use Trans to build your agent and you run the agent on Agent Core. But the way we built Agent Core, you can mix and match. Let's say you want to use some other gateway, you can use that gateway with the Agent Core runtime and the tooling around, this is just going to get better and better and better.

So does Agent Core. Is that like the GPUs where the inference is happening or it's one level higher?

That's bedrock. Agent Core is one level higher than that. The way to think about agent core runtime is it's lambda for agents. That's the simplest definition I can give you. It's a serverless runtime, but you're not putting functions on it. You're running an agent.

Gotcha.

Okay, cool.

That makes sense. Now, I'm sort of wrapping up here, but there's a few questions I'd love your take on. When it comes to, like, the whole AI agents might take our jobs, right? That's obviously a concern that people have. What's your take on it? What sort of advice do you give people that have those concerns?

I'm going to speak from a software development perspective because that's the world I live in. The better and more you are as a software engineer, the more effective you get at using these agents. As a software developer, you get stuck all the time. And when you're using agents, you can get yourself unstuck much faster. There are problems that you've not wanted to solve or you haven't had time to solve that you're able to get unblocked on so that you can solve the really hard ones. The skill that I tell software developers to think about, and you can see this because you can see the ones who are good at that, be very effective. At Amazon, we have a principal engineering tenet called Illuminate and Clarify. Our most senior engineers are really good at it. They're able to illuminate a problem and clarify it, simplify it, break it up into smaller things. And any software developer, and this applies to any domain. As you look at a problem, how are you able to plane it? How are you able to express it in natural language, potentially break it up into smaller chunks and give that context to the LLM, to the agent, the more effective you're going to be. How do you interact with agents is a little bit different than how you may have interacted with an IDE for writing software in the past, but once you figure that out, the outcome is much, much better. That's what gets people excited. So you have to change the way you think. You're type thinking in natural language. You're thinking in terms of breaking up problems. In fact, the whole goal of KIRO was how can we give you a user experience that actually makes it easy for you to do that? How can we help you think like that and then express them by. That's how your interface works. But the idea is that the person at the other end of it, the more. If you're a really good developer who can think like that, you're going to be even more effective. That's very much the approach we took. What you're not doing is typing code yourself, but that's okay. That's busy work. That's typing, right?

Right, yeah. And I mean, that sort of answered the next question I had lined up, which was going to be, what advice would you have for, like, coders and engineers? Do you think it's still a valuable path to study?

Absolutely. I mean, we see people being so effective, and over 80% of the software developers at Amazon are using AI tools for their work. And the thing that I tell them, or anybody who asks me, is it's less about how well, you know, JavaScript or Ruby or Rust or Python, it's more about how you think about breaking a problem down, giving a goal to an LLM, understanding how to express that goal, and then having the skill and judgment to understand, to help it move along. If it's going in the wrong path, how do you tell it? Hang on. A lot of what my team is trying to do and what AWS is trying to do is give you the tooling that helps you do that more effectively.

Yeah, that makes sense. I actually had a conversation recently with Kevin Scott, who's the CTO over at Microsoft. And during that conversation, he actually mentioned that he believes that the rise of AI is creating more coders, not less, and that people starting to use, like, these various coding tools, they're actually starting to make progress, which is getting them excited about code. And when they run into bugs, they go and want to learn more code so that they can better troubleshoot their vibe. Coding. Right. And he actually believes that we're going to see a rise in people wanting to become coders as a result of AI, not the other way around, and happier coders.

I'll give you a very concrete example. You've probably heard this one before. Andy's talked about it. We built a transformation agent that takes Java code from one version to the next version. Right. And in the history of any software company, when you have to do version upgrade, it's a pain. And pain because you have to go package by package. Every team has to do it. And there's. When you have millions of packages, it takes a long time, like years sometimes, because it's work that nobody wants to do.

Right.

So what we did was we took our transformation agent and plugged it into our build systems and how we do work at Amazon, and a small team was able to transform very large, hundreds of thousands of packages. If I Remember correctly. And what they did was they shipped code reviews to the teams that owned the package. The teams themselves didn't need to do anything other than code reviews. And the outcome was it saved us 4,500 developer years and $260 million in ARR from the infrastructure savings. The interesting point is now suddenly you had the people who originally would have been spending their time doing the upgrades, actually writing software and building, creating and innovating, which they would much rather do. Right. And I think that's what speaks to why, as a developer, you're going to get more excited and there'll be more because you can accomplish more.

Absolutely. So, you know, I think this is probably the deepest we've ever gone on the show into AI agents. And so, like, I think this really sort of gives people a pretty clear picture of what AI agents are right now and where they're all headed. So I really, really appreciate you hanging out and taking the time with me. Is there some sort of resource or URL or somewhere we should send people after listening to this episode to go learn more about what you guys are up to?

Yeah. So for my team, I would recommend people go to kiro.dev that's a website for Kiro. It gives you a good idea of how we think about software development in agent world. And if you want to learn about Agent Core and other agentic stuff that we are doing at aws, go to the AWS website and search for agentic AI. You'll find all the stuff that we do there. And you know, we love feedback for kiro. For example, we have a Discord Channel. Join us there, use it, tell us what you'd like to see. Our goal is to make it easier for you to build and run your agents and use them.

Amazing. Well, Deepak, this has been an absolutely fascinating conversation. I've had a blast chatting with you. So thank you so much for hanging out with me today and going deep on agents with me.

Yeah, thanks for having me.

And that wraps up today's episode of the Next Wave podcast. A couple of big takeaways from this conversation with Deepak. First, AI agents aren't just chatbots with a fancy name. They're systems with autonomy, reflection, and the ability to actually solve problems on their own. We heard real world examples from Formula One teams cutting analysis time by 86% to Genentech saving five years of manual research work, to Amazon saving 4,500 developer years and over $260 million by using agents for code transformation. So the big lesson here Whether you're running a startup or a Fortune 500, AI agents are about to fundamentally change how work gets done, and now is.

The time to start experimenting with them.

I want to thank Deepak again for joining me and for sharing such a clear inside look at how AWS and Amazon are building the agentic future. And to everyone listening, if you enjoyed this episode, make sure to subscribe to the Next wave podcast on YouTube, Spotify, Apple Podcasts, or wherever you like to listen. That way you'll never miss the conversations that help you stay ahead of what's coming next. Thanks again for tuning in. I really appreciate you. Hopefully we'll see you in the next one.

Bye Bye.

Sat.