In this interview during theCUBE's coverage of AWS re:Invent, Clare Liguorui, senior principal engineer at AWS, joins theCUBE’s John Furrier to detail the rapid adoption and evolution of Strands, an open-source framework for building AI agents that originated internally at Amazon. Liguorui reveals that Strands has surpassed 3 million downloads since its mid-year release, a testament to the industry's shift toward a "model-driven" development approach. She explains how this philosophy minimizes boilerplate code by allowing frontier models to handle reasoning and tool selection, rather than relying on brittle, hard-coded workflows. The discussion highlights major announcements for the framework, specifically the introduction of a TypeScript SDK and support for edge runtimes, which broaden accessibility to the massive JavaScript developer community and enable agents to run anywhere from browsers to factory floors.
The conversation delves into the practicalities of deploying agents at the edge, where Liguorui describes the necessity of hybrid architectures that combine fast, local models for immediate tasks – like robotic motor control – with powerful cloud-based frontier models for long-range planning. She also introduces the concept of "Strands steering," a mechanism designed to keep agents aligned with specific operating procedures (SOPs) through reminders, acting as a guardrail against model drift during complex, non-deterministic interactions. Liguorui emphasizes that these advancements allow developers to treat AI agents not just as experimental code, but as reliable, production-ready assets that can integrate seamlessly with AWS services like Lambda, Bedrock and SageMaker while remaining flexible enough to run on a laptop.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
AWS re:Invent 2025. If you don’t think you received an email check your
spam folder.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open the link to automatically sign into the site.
Register for AWS re:Invent 2025
Please fill out the information below. You will receive an email with a verification link confirming your registration. Click the link to automatically sign into the site.
You’re almost there!
We just sent you a verification email. Please click the verification button in the email. Once your email address is verified, you will have full access to all event content for AWS re:Invent 2025.
I want my badge and interests to be visible to all attendees.
Checking this box will display your presense on the attendees list, view your profile and allow other attendees to contact you via 1-1 chat. Read the Privacy Policy. At any time, you can choose to disable this preference.
Select your Interests!
add
Upload your photo
Uploading..
OR
Connect via Twitter
Connect via Linkedin
EDIT PASSWORD
Share
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
AWS re:Invent 2025. If you don’t think you received an email check your
spam folder.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open the link to automatically sign into the site.
Sign in to gain access to AWS re:Invent 2025
Please sign in with LinkedIn to continue to AWS re:Invent 2025. Signing in with LinkedIn ensures a professional environment.
Are you sure you want to remove access rights for this user?
Details
Manage Access
email address
Community Invitation
Clare Liguori, AWS
In this interview during theCUBE's coverage of AWS re:Invent, Julia White, chief marketing officer of AWS, joins theCUBE’s John Furrier to break down the major announcements reshaping the enterprise AI landscape. White details the launch of "AI Factories," an opinionated infrastructure approach designed to bring large-scale compute capabilities directly to customers' existing data centers, specifically for highly regulated and sovereign needs. She explains how the new Nova Forge empowers organizations to create custom frontier models by securely blending their proprietary data with Amazon’s training data, effectively solving the trade-offs between fine-tuning and model performance. The conversation also highlights significant advances in custom silicon, with new generations of Trainium delivering up to 80% better price-performance to optimize AI infrastructure from top to bottom.
The discussion then shifts to the emergence of "Frontier Agents," a new class of autonomous, massively scalable and long-running AI agents capable of executing ambiguous tasks over weeks without constant redirection. White outlines AWS' strategy to democratize agentic AI through AgentCore, which provides the essential enterprise building blocks – such as security, governance and identity management – needed to move agents from experimental fringes to production environments. She emphasizes how these innovations, alongside specific agents for software development, DevOps and security, are unlocking faster time-to-value and fundamentally changing how businesses approach the software development lifecycle.
In this interview during theCUBE's coverage of AWS re:Invent, Clare Liguori, senior principal engineer at AWS, joins theCUBE’s John Furrier to detail the rapid adoption and evolution of Strands, an open-source framework for building AI agents that originated internally at Amazon. Liguori reveals that Strands has surpassed 3 million downloads since its mid-year release, a testament to the industry's shift toward a "model-driven" development approach. She explains how this philosophy minimizes boilerplate code by allowing frontier models to handle reasoning and...Read more
exploreKeep Exploring
What is the design philosophy behind Strands and how do they relate to frontier agents?add
What are the platforms and environments where Strands can be run, and what benefits do they provide?add
What are the capabilities of the Strands tool regarding data formatting and access?add
>> Hello, I'm John Furrier with theCUBE, here in Seattle at the AWS headquarters at re:Invent, getting all the action on re:Invent news and all the big announcements on stage. Clare Liguori is here, senior principal engineer at AWS, talking about Strands. Great to see you. Thanks for coming on theCUBE and special news analysis on stage and the big reveals. Thanks for coming on.
Clare Liguori
>> Thanks. Happy to be here.
John Furrier
>> This is our 12th year covering re:Invent and every year is great, but this year seems a little bit more game-changing in the sense that the AI game is on. Everyone sees it. It's the top line, all the boardrooms, all the engineers are working on it, people coming out of retirement because with coding assistants and Kiro, all the cool things, it's very easy to do stuff now at scale. Plus, the infrastructure investment is inline with the pace of the demand. So, you're starting to see the formation of the agent software stack model. You're starting to see the formation of startups that are AI-native now. It was cloud-native, they're still around, they're on the cloud, but cloud and AI native are converging. And Strands is one of the most exciting things we've covered because it grew organically here at AWS. And now, it's like becoming the key linkage and coordination, connective tissue, whatever we call it, but with the models. So, tell us why the news is so big?
Clare Liguori
>> Yeah, that's right. Strands grew organically out of a need that we had in Amazon for a simple way to build agents. And we decided we had so much success with it internally that we decided to open source it for our customers. And to this re:Invent, we are announcing two big expansions to that. One is TypeScript SDK support, in addition to Python. And the other is support for edge runtimes with Strands.
John Furrier
>> And the uptake on terms of the developers, what's been some of the stats? You guys had like millions of downloads. What was the number?
Clare Liguori
>> We've had over three million downloads of Strands since its release, middle of the year.
John Furrier
>> Okay. So, we're also heard about the whole frontier models, Nova opening up, democratizing. Frontier agents opens up a whole nother level of value because now you can take the coolness of agents and make it operational with all the stability. How does the Strands fit into all this? How would someone think about Strands for, say, frontier agents?
Clare Liguori
>> We have a design philosophy, really two design philosophies. One is as easy as possible. We want you to write as little boilerplate code as possible, just a few lines of code to get started writing a real agent. The other is what we call the model-driven approach, which is instead of writing a lot of boilerplate and a lot of logic around the models, let the model do the driving. We've found that frontier models especially are so capable of reasoning and driving the tool selection and things like that in the agents that the more that you do almost around it, the worse your agent gets is what we actually found internally. And so, simplifying it and pairing it back to as simple as possible and integration with the model means that you're able to take advantage of the full capabilities of the model. And then, as the models are getting better, which we're seeing every few months, you don't have to do anything to your agent other than change which model you're using for your agent to suddenly get better. You don't have to change your entire software stack around it.
John Furrier
>> So, that brings up what we've been seeing the progression of how people have been playing with models. They go, "Oh, I grabbed this model." Then they go down the road and the better model comes out, then they have to walk that back and then come in. You're laying out this optionality play, where you create software around models that may be available and you can almost pick the right model for certain use cases. Is that right? Am I getting that right?
Clare Liguori
>> That's right. You can think about cost, you can think about complexity of the tasks that you're doing. You can think about particular use cases that the model might be better at or not and make that selection and change it down the road. Often I find what happens is that when we're prototyping an idea for a new agent, you pick the best model possible, right? You pick the absolute best of the best to just prototype and think through, is this even a feasible idea? And then you start to get thinking about production and cost and efficiency and you start to optimize, but you don't have to change anything about your agent itself. It's just swapping out the model and seeing how it performs.
John Furrier
>> It's like that scene in The Matrix where you can just upload the best model for the task. Matrix, if you've seen that movie. Maybe the younger generation might not have seen The Matrix, go check it out. But this brings up the model-driven approach. I do like this because we've been seeing it in some of our observation space with theCUBE and SiliconANGLE, young startups, they're not thinking, "Where do I host the servers?" They're thinking, "How do I build within the models themselves?" So, they're optimizing for their end game, which is the app goal, which is the outcome, but they're using the models first and then writing code, distilling. And so, now you have Nova. So, you have with that new approach, Nova with the open data, this is a different development innovation formula. How would you describe this? Because this feels like very much the AI-native side. Now once they do that, they still got to run it on AWS infrastructure. So, how legit is that? Do you see the same thing? And how would you frame this new trend?
Clare Liguori
>> One thing that we are starting to see is that thinking about development is very different with agents. So, normal traditional programming paradigms, you write a sequence of steps and you might turn those into a workflow. There might be a loop somewhere or an if statement condition somewhere. Agents are completely different because they have such reasoning capabilities. Often we're finding that for complex tasks in the past when models weren't as capable, people would go back to those same programming paradigms that they're familiar with, create a workflow. But as it turns out, that's a very brittle approach. You have to think about all the edge conditions. Your agent isn't flexible for unexpected use cases that might come in. And so, we found that when we did this inside Amazon, we were spending so much time trying to get these agents to do what we wanted them to do. And when we went back to the model-driven approach, we found that we can actually guide and steer the agent and let the model come up with the workflow itself and still have the capability to do really complex tasks in the agent. So, it's a very different mindset among younger companies, younger developers that are coming up in this completely non-deterministic AI world.
John Furrier
>> And I think that's what's nice about this agent world is that there's a lot of value in that non-deterministic way because that's the goal of the AI to be reasoning. You brought up the edge, we'll get that in a second, but the TypeScript, you said it's one of the most popular languages and you mentioned Python. Python, some would say, is the most popular. So, talk about the difference between TypeScript and Python. Why agents like TypeScript over say Python?
Clare Liguori
>> Yeah. I'm really excited about Typescript SDK being in preview for Strands because Typescript is based on JavaScript, and the two combined are the most popular programming language ecosystem that there is. And so, we're really broadening access to more and more developers who are used to working in Typescript and JavaScript. But the other part of it is now the agents can run in new places. They can run in the browser, they can run anywhere that JavaScript can run, which is pretty much anywhere. And one of the things that I've noticed is that it's potentially a lot easier to write your agent because you get a lot of the nice characteristics like type safety and TypeScript that you don't get in Python. So, you're actually able to almost build your agent faster because you get compile time errors and things like that. Versus Python, you may not know there's a problem until you actually run it.
John Furrier
>> So, it's like letting it run out in the wild and then it breaks, then troubleshoot... So, there's better safety you're saying with TypeScript?
Clare Liguori
>> That's right.
John Furrier
>> All right. Let's talk about the edge because the edge is something that you mentioned. We've been covering AI factories now for over a year. This whole theme is one of our popular programs. In the series is robotics. Robotics and AI are going hand-in-hand, not obvious, but it is obvious to think about it because hardware and software are merging. So, talk about why the edge is important because robots need to have some brains too or devices or even any end user experience is going to need a model.
Clare Liguori
>> Yeah. There's two parts to the edge capabilities that we are announcing. One is bidirectional streaming support. So, this enables real time communication between both agent to agent and human to agent. And then, the other part is going to be more local streaming support. So, if you think about a robot and a factory, they can't spend milliseconds waiting for a call to a frontier model in the cloud in order to move the robot arm. So, we've seen that our customers in these industrial manufacturing, autonomous vehicles, all of these industries are increasingly using local models, of course, but they tend to be small. They tend to be not as capable as these frontier models for reasoning and long-range planning. So, you get the best of both worlds with this launch because you can run a local model down in the factory floor with Strands. And then, you can also run Strands in the cloud where you can access these long range planning capabilities in the frontier models and combine the two together.
John Furrier
>> Yeah, Clare, this comes up a lot around robots because they're getting so much stronger and intelligent, whether they can get more precision on gripping. We're seeing all kinds of use cases, not just the humanoids, everyone thinks humanoids, but just like manufacturing, construction, all these use cases where robotics are really doing well because of the software. So, I have to ask, on the model speed, what is the performance threshold? You mentioned milliseconds, how do you guys view that? Because I mean, you're driving a car, you can't miss... If it's a self-driving autonomous vehicle or a robo-taxi, everyone sees as an example. But that as that goes to physical world in the edge, retail, manufacturing, what are some of the performance things that the models need to pay attention to and how does Strands help that?
Clare Liguori
>> Well, we tend to see two different models at play here. One is called VLMs, visual language models, and that's the model that can translate what is this robot sensing, with usually a video stream coming in? And how does that turn into what is it seeing? Is it seeing an apple? Is it seeing a car? And then the large language models are really turning that language of all of these inputs that are coming from the real visual world into a plan of action. And then, that's going to get turned back into individual motor controls for moving the arms. And so, you tend to have these very specialized models for those different tasks locally versus this frontier model in the cloud.
John Furrier
>> And processing power becomes key and networking are key elements in the edge?
Clare Liguori
>> That's right.
John Furrier
>> All right. Talk about reminders. This came up, you guys are doing a reminders. What is that all about?
Clare Liguori
>> So, we're calling that strand steering, but it's effectively a reminder to the agent. So, one of the things with the model-driven approach that we've found is that often people want to guide the agent a little bit more in production. As you know, you might test it a couple of times on your laptop, you might test it a few more times before it goes to production, but now you've got thousands of people trying it out. How can you be sure that it's actually following the behavior that you want to give it? One is Agent SOPs, which we just open sourced on GitHub, which is a technique that we've been using inside of Amazon to automate work. Agent SOPs is basically a way to structure your prompt to describe the way to accomplish a task. But often, as the conversation gets longer and longer with the agent, we need to give it a little reminder. We need to give it some more steering and ensure that it's staying on task. And so, reminders and steering providers, it could be an LLM prompt, it could be just simple code that's validating the agent's behavior. And then, you give a little prompt back to the agent that says, "Hey, you need to be asking about this now," or, "Hey, you should not be doing that task right now."
John Furrier
>> So, it literally is almost like keeping it in its lane?
Clare Liguori
>> That's right.
John Furrier
>> It's like stay in your lane?
Clare Liguori
>> Yeah.
John Furrier
>> That's interesting. And now is that more of a reset because most people think about model drift, and then you have to re-prompt. Is that the kind of same concept?
Clare Liguori
>> It's more steering it. As you see it drift off, get it to go steer in the other way. And you can hook this into really key lifecycle events. You can hook it into a tool call or a user input response in order to look at, are we still on track with that same agent SOP that I want you to follow? So, one example here is something like you want to build a chatbot that does travel booking and you want it to gather a certain amount of information from the user, right? Where do you want to go? What day? Often we'll find that a workflow won't work here because it's a conversation. It might go in a lot of different orders of how you collect that information. The user might give you all of the information upfront or they might wait for you to prompt to ask them. And workflows in this case tend to be very brittle for that specific use case because it's a conversation, it could go a million different ways, but steering is able to give the agent a little reminder that says you're supposed to ask about date, you're supposed to ask about destination before you start go and actually booking something. So, it's fascinating-
John Furrier
>> It's almost like QA process for the agent? Stay on task and be calibrated?
Clare Liguori
>> Mm-hmm. That's right. Some good feedback for it.
John Furrier
>> So, the Typescript thing is huge. Looking at the JavaScript, that piece of it, that's going to be a big developer in. Python already is huge there too. Looking at the big picture with TypeScript support SDK, what's that going to do for the developer? How do you see it fitting in? Does it fit into frontier agents? Where in the Amazon higher-level services does it fit? I mean, look at serverless, that came out of internal work and that just created a whole nother generational shift, Strands. Where does Strands fit into the picture?
Clare Liguori
>> Strands can be run in Lambda serverless. It can also be run in Bedrock AgentCore runtime. It can take advantage of all of the models within Bedrock, including your own foundational models that you host there. It can take advantage of models that you're hosting in SageMaker. So, it has all of those enterprise integrations into the AWS services. But one of the things that I like is that you can also run it anywhere. You can run it on your laptop for testing. Now, with the TypeScript SDK, you can run it on your website if you want to. So, we're really enabling you to run it anywhere that you need to, along with the power that AWS brings with our AI services.
John Furrier
>> So, it's going to bring a lot of user experience, visual, presentation layer capabilities for agents to play with the interactions with users?
Clare Liguori
>> Right.
John Furrier
>> And also make the memes better. They're already crazy right now with the AI. I mean, this is where starting to see the fun kick in. For developers, what's the pitch for them because this sounds like it's worth investing in? What would you say to the engineers out there and the developers around how to think about Strands, how they should think about the long game of where it goes? What would be your-
Clare Liguori
>> Well, getting started is very, very easy. We wanted to make it so easy for anybody who is not an AI expert to be able to write an agent in a few lines of code. I've had product managers come to me and say, "I wrote a Strands agent and this is amazing." So, not just-
John Furrier
>> I got five ideas already I'm going to use it for.
Clare Liguori
>> Exactly. And it's really easy to use with your coding assistant. We have tools for your coding assistant that can make it really easy for it to generate an agent for you, so very easy to get started, but we're also doing a lot of experimentation and research on the bounds of agents and pushing the model-driven approach. So, the experimental steering capabilities is an example of that where we're starting to look at how can we push to more and more complex tasks, so that it really scales from your first agent to everything that you might want to do in production with agents.
John Furrier
>> Yeah. In the model-driven approach, you guys have to be sensitive to, okay, let's test the model. Does it fit the parameters of the work of the agent and the tooling around it?
Clare Liguori
>> And what we've seen is that actually a lot of models now, including some of the open-source smaller ones that are coming out, are very capable for reasoning and planning with Strands. And that's just a drop-in replacement because we have all of this model support already. So, I've been playing with gpt-oss-120b on Bedrock, playing with some of the Qwen models. So, all of those newer models on Bedrock we already support on day one.
John Furrier
>> And with the Nova news, they're going to offer the Amazon proprietary data and allow people to bring their data estate in, have their own frontier model. How does that fit in? Because that's going to be a huge enabler on the enterprise side and startups where, basically for cloud-based pricing, you can have your own frontier model with your data?
Clare Liguori
>> Yeah. Bringing your data is so important for agents. Today, all that we have is, for most of us, not fine-tuning models, what we have is effectively access to databases and access to knowledge bases. But often, you find there's a misalignment between what the model understands from its training data versus the data that you have, the concepts that you have.
John Furrier
>> Is there a format that has to be compatible with Strands? Like for example, does everything have to be a vector embed or can it just use on all unstructured-structured data, or does it have to be formatted?
Clare Liguori
>> It doesn't have to be formatted. It can be structured or unstructured. The tool capabilities in Strands is very flexible. So, in Python and in TypeScript, you just write a function and you can do anything you want. You can return structured data, unstructured data, whatever database you need. There's also community tools based around some of the really popular vector databases out there, as well as Bedrock knowledge bases. So, all of that information can flow right in. We've got S3 vector support at this point, so you can just put your data in S3 and search it there. So, it's easier than ever to access your data from your agents.
John Furrier
>> Clare, thanks for sharing and unpacking the Strands announcement. Thanks for coming on.
Clare Liguori
>> Thanks.
John Furrier
>> Appreciate it. I'm John Furrier. Getting the news and getting a deep dive on what it means and what it means for the future. And again, the developer scenario here is continuing to get better and better with AI. And the technology that's enabling to write the AI-native apps, model-driven is a great approach. TypeScript's coming in, making it easier for the end user. A lot more value coming. Of course, we're doing our part. Thanks for watching.
John Furrier