🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - AI agents in action: Architecting the future of applications (INV202)
In this video, AWS Director of Technology Shaown Nandi explores the agentic AI revolution, demonstrating how agents deliver exponential business value beyond incremental gains. He explains how agentic patterns differ from traditional architectures by moving logic from static code into LLM-augmented agents capable of dynamic reasoning. AWS introduces Strands Agents (open source) and Amazon Bedrock Agent Core to simplify building production-ready agents. Customer speakers share impressive results: Stedi built healthcare eligibility troubleshooting agents in two weeks achieving 33% resolution rates, QAD achieved 26% productivity improvements and 30% inventory cost reductions in manufacturing, and Pattern increased e-commerce conversion rates by 2% while reducing photoshoot time by 96%. The session emphasizes four key advantages of agents: accelerated execution through parallel processing, adaptive orchestration with intent-based reasoning, embedded quality assurance, and persistent memory that compounds learning over time. Nandi positions this as a major inflection point comparable to cloud adoption, where agents will become foundational to every business process.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Welcome to re:Invent: From Cloud Skepticism to the Dawn of the Agentic Era
Please welcome to the stage Director of Technology for AWS, Shaown Nandi.
Hi everyone, welcome to Reinvent. Is everyone having a good start to their week? I am still blown away by Matt doing 25 announcements in 10 minutes, hitting that shot clock. This is actually my 10th year at Reinvent, and it feels like a bit of a full circle moment for me. Twelve years ago, I watched a video of Reinvent and saw the CIO of Dow Jones stand on stage and say that they were planning to move 75 percent of all of their workloads to the cloud. This was back in 2013, and I remember thinking he was crazy. It's not possible. You can't move HR systems to the cloud. You can't run financial systems in Europe in the cloud. There's just no way.
Skip ahead two years later to my first Reinvent in 2015, and I was then CIO of Dow Jones. I started to see what was truly possible. There were fewer than 20,000 attendees. Everyone at Reinvent was just in one place right here in the convention center. It was just one hotel back then. There was a buzz, and I was hearing about all the workloads that everyone was moving. Now, for the last five years and more, we can see that every workload can run in the cloud. That's not a constraint anymore.
Today, I see a lot of parallels from 10 years ago. Today we're living through the dawn of the agentic era, and I really feel like when we go back and think about 2025, it will be clear that now is when business really changed. The AI revolution is really happening. Why am I so optimistic about agentic AI? Well, that video we opened with gives you a little bit of an idea. We're not all the way there yet, but if you squint just a little bit, you can see that fast approaching future where companies have hundreds or maybe thousands, or perhaps tens of thousands of agents performing really complex work and driving really significant value. I call this the agentic advantage, and it's huge.
The Agentic Advantage: Exponential Value Through AI Automation
Because when we talk about results, we're not thinking about single digit gains because some chatbot was bolted onto an existing business process. We're talking about exponential value. The type of value you don't just get from a little bit of optimization or a little bit of cost saving or a little bit of efficiency. Let me give you just a few examples. So here at Amazon, we've been harnessing agents to drive greater automation and efficiency. Earlier today, for those who attended the Amazon innovation talk, my colleague Dave Treadwell, who runs all of Amazon's e-commerce foundation, shared that the Amazon store's business unit now delivers more than two billion dollars in annual cost savings from AI automation and defect elimination. Our CEO has talked about how we used agents to modernize Java applications, saving 4,500 years of development time in just one set of upgrades. Not annually, just one set of upgrades done across tens of thousands of systems. By adopting an AI-native approach to software development, Amazon has seen the productivity of our development teams increase by a factor of 4.5.
That is incredible results. But it's not just Amazon. Customers that have embraced agentic AI are also seeing exponential value. Customer service company ASAPP automated 90% of their end user interactions using agents. They increased their first call resolution rate to 91%, while reducing costs by 77%. Better outcomes and lower cost, a rare combination. You see the same thing at Systems Integrator NOVACOMP. Who use agents to modernize applications, they saw a 60% reduction in technical debt, perhaps some of that stuff that was addressed earlier today, and were able to modernize instances, a task that used to take them 3 weeks, they were able to do in just under an hour.
And Brazilian healthcare nonprofit, noharm.ai, was able to boost patient capacity by a factor of 8, while saving more than $30 million per year across the hospitals they serve. These numbers are pretty incredible. And they're driven by agentic AI. But why? Why do agents make such a difference? And what's fundamentally different about architecting agentic applications? What is it that enables this exponential value? It's fundamentally two things.
Why Agents Make Such a Difference: New Patterns and Business Value Calculus
First, agentic patterns are enabling business value in a way that has been difficult to do with generative AI up until now. When we first started, generative AI applications had a pretty constrained set of architectures that required a ton of engineering effort to build. Fast forward to today, agentic approaches are much more flexible with distributed system architecture. There's a lot less engineering complexity needed to have an LLM and an agent be integrated right into your business processes.
Second, these new agentic patterns are translating to a much different business value calculus than we've had before. With generative AI there was actually a decently high barrier to entry. Integrating your data might require a lot of engineering work. So the use cases had to deliver a really good ROI. The juice had to be worth the squeeze. Today, with the evolution of what we've seen in the last year, the barrier to entry has gone way down. Whether you want to automate an entire workflow or do a lot more targeted investment, both options just don't require the same amount of time, money, or effort.
And the use cases have changed. Before agents, value had to come from a simple prompt and response workflow. And the outputs of those workflows had to be differentiating enough to drive really outsized value. They had an impact, but the impact was often incremental. And it was usually limited by a human's ability to go back and forth with an LLM. Now the game's changed. The flexibility in where the automation sits and how the LLMs are incorporated in their workflows is completely different. And the value delivered is exponential because agents are doing things a human being can't or won't do in the same time and capacity.
Architectural Evolution: From Cloud Native to Agentic AI Patterns
These two pillars—flexible agentic AI patterns and the new math around business value that agents can unlock—set the stage for today's talk. And we have some bleeding edge customers joining us. Three CEOs from Steady, QAD and Pattern, who are going to share their agentic journeys. Let's dive right in. First, the architecture patterns themselves. This is where the biggest shift is happening. Traditional serverless and event-driven architectures have absolutely inspired agentic AI design. They gave us all the foundations: elasticity, scalability, and the ability to decompose complex applications into independently deployed services. These principles are just as important today as they were when we first moved to cloud.
What's fundamentally different now is where the logic of a system lives. In a traditional application, we all know this, the core logic is static. It's written in code. Every conditional branch, every business rule, every exception has to be explicitly programmed by a developer. In an agentic system, that logic moves into an LLM-augmented agent. The intelligent components that can interpret context, make decisions, and act toward a goal with defined boundaries. They reason around which tools, which data to leverage, they can adapt their behavior dynamically as inputs change. That shift might sound subtle, but architecturally, it's profound. We're moving from deterministic systems where outcomes are predefined to systems capable of dynamic interpretation and intelligent augmentation. And we've been on this journey for a while.
Cloud native architecture patterns have been the cornerstone best practices we've all learned over the last two decades. You're all experts at it. They played huge roles in shaping the development of the cloud and how companies run today.
Then generative AI appeared. Early on, organizations tried to harness the LLMs, but the architectures were often rigid and isolated. You typically had a single model endpoint that could summarize text or draft a response, but it was really difficult to integrate that into an existing production system in a way that respected cloud native design principles. Tool integration was hard. Access to enterprise data was hard. Scaling those workloads was hard.
The model interaction tended to be stateless and incremental. One prompt, one response out. There was no concept of sustained context or ongoing goals. Even though the potential was enormous, we all saw that movie trailer and got really excited. It felt difficult to build generative AI applications that looked and behaved like modern distributed systems, those concepts we spent decades perfecting. Now you could do it, we have lots of success stories, but it came with engineering complexity. This is why so many companies had early POCs that stalled before reaching production.
Today, agentic approaches have changed the equation completely. Because models can maintain goal-driven loops, they stay focused on a complex objective across multiple steps and iterations. They retain context and memory, integrate tools natively, and operate within scalable event-driven environments. Runtime capabilities like Bedrock Agent Core now provide a simpler, more operationally robust foundation for deploying and supervising agents in a consistent way.
You can define capabilities, enforce policies, monitor behaviors, and let agents coordinate through emerging protocols, rather than building your own custom APIs. These innovations have relaxed many of the constraints that made those original generative AI architectures seem so brittle. In fact, if you look closely and you're using the right tooling, there isn't a big difference anymore between cloud native patterns and agentic patterns. What's really happening is the same foundational practices that I've been using to drive architectural success for decades are now being turbocharged by placing an LLM at the center of the system. That's what's unlocking completely new levels of value creation.
Building Production-Ready Agents with Strands and Amazon Bedrock Agent Core
Now, you can understand the fundamentals of these new agentic patterns. The concepts of memory, collaboration, and goal-driven orchestration. It's easy to see the potential, but actually building secure, observable, and production-ready agents is a different story. Until recently it required lots of custom plumbing for session handling, tool registries, memory stores, observability, permissions, and so on.
That's why we built Strands Agents and Amazon Bedrock Agent Core to take the undifferentiated engineering lift off your plates. Now with Strands, you can focus on defining exactly what you want your agents to accomplish, not how to orchestrate them. Strands handles the goal-driven loop, the planning, acting, and reflecting cycles that connect your prompts, models, tools, and data services. It abstracts away the non-functional plumbing that every enterprise needs. Like security, observability, and deployment pipelines. So you can move from prototype to production with far less complexity. We've chosen to open source Strands and make it fully available to all of you.
Now, when it comes to the challenge of running those agents, even if you built the powerful orchestration layer, operating at scale, especially in an enterprise context, can be difficult and potentially risky because autonomous systems introduce new reliability and safety considerations. That's where Amazon Bedrock Agent Core comes in. Agent Core provides a managed enterprise-grade runtime platform for agents, covering IAM, session isolation, long-running workloads up to eight hours, tool integration, memory services, support for any model and framework, and support for policy and support for evaluations as of this morning.
It enforces policies at scale securely. You have the flexibility to compose advanced agentic systems, but you can do it with the reliability, governance, and performance you always expect from AWS. Together, Strands and Agent Core remove the friction that we've all been fighting, transforming what used to take months of engineering into a set of fully manageable, composable services that you can build within hours.
They let you think in the same system-level terms that all of you already know, but supercharged with reasoning and autonomy built in. This is where things are today—what we're experiencing ourselves at Amazon and what we're hearing from customers. But the space is changing rapidly. A lot of what's possible right now wasn't possible at the beginning of the year, and 12 months from now, this landscape is going to look different again. That's what's exciting.
Customer Spotlight: Stedi and QAD's Agentic Journeys in Healthcare and Manufacturing
As we continue to evolve, AWS will be here to share our learnings and experience, our approaches, and best practices to help your agentic journey. Two customers that are really far along on their own agentic journey are healthcare technology leader Stedi and manufacturing technology leader QAD. To hear more about how they're building with agents today, please welcome founder and CEO of Stedi Zack Kanter, CEO of QAD Sanjay Brahmawar, and AWS Director of Technology Olawale Oladehim.
Thank you so much. I'm very excited to get into this discussion, so thank you both for making the time. I would love to start by hearing a little bit about your companies and how you're thinking about agentic technology and highlight your technologies and architectures. Zack, I'll start with you.
Thanks for having me. I'm really excited about all the launches and everything so far. Stedi is a healthcare clearinghouse, so we process transactions between two parties in the healthcare ecosystem: providers and payers. On the provider side, that could be your local doctor or dentist, or at large scale, it could be a large academic medical center or hospital in a big city. On the payer side, you're dealing with insurance companies like Blue Cross Blue Shield, UnitedHealthcare, and Cigna, but there's also a long tail of thousands of regional payers and third-party administrators that we connect with.
To give you a tangible example, these are all related to insurance. You go into your doctor's office, they ask for your insurance card, they tell you your copay is $20 and you have 3 visits left. That is something called a real-time eligibility check, and that's processed on Stedi's rails. That's the happy path when you have your valid insurance with you. When you think about the unhappy path, maybe you forgot your insurance card, or it's smudged out, or your insurance has changed since the last time you got your insurance card, or worst case, maybe you came in an ambulance and don't even have your wallet with you.
In those situations, you're relying on somebody's memory and then transcription—human transcription or typing—in order to get those details right. If your name is Bill or Katie, but your name is actually William or Katherine, or it's Katherine with a C instead of Katherine with a K, or if your name has been misspelled over the years, you run into this problem. Insurance companies are very brittle by design. They don't want to return PHI for the wrong person, so if the details are wrong, they reject the insurance check.
The problem is that when that gets rejected, phone calls have to be made. Somebody picks up the phone, or worst case, they tell the patient that they're not covered. We publish content about how to troubleshoot these things—blog posts, documentation—but the reality is that people are pretty busy and can't implement all the best practices. What we built as our first agentic feature, we have a saying: no restaurant ever went out of business for being too small. We picked the narrowest use case, which is troubleshooting these eligibility checks.
We built it using Amazon Bedrock AgentCore and Strands Agents, and we built it end to end in two weeks, which I think is pretty quick. It's a testament to a couple of things. We scoped the problem down a lot, and the technologies that were provided to us by AWS were fantastic. When you look at AgentCore and Strands, it's exactly what you'd hope for in the primitives you're looking for. We get to use all of our familiar tools to build our MCP server, like Lambda, API Gateway, deployment pipelines, and all the things we have come to love.
That's fantastic. Sanjay, I'd love to hear a little bit about QAD's use case and how you use agents. Well, first of all, thank you for having me, and I have to say it's absolutely awesome to be here. I thought the keynote was fantastic—the amount of announcements and AWS rocks. You guys move so fast. In manufacturing, we say speed is no longer a strategy; it's basically survival. AI needs to be real, and it's here.
I'm the new CEO of QAD, just started in March this year. QAD is about half a billion in size with about 1,500 talented people.
We are a global software company based out of Miami, and we are the number one AI-powered manufacturing platform. We are effectively turning mid-market manufacturers into champions of manufacturing. When I say mid-market, I mean companies with half a billion to about 55 billion in revenue. Our mission is to take manufacturing from systems of record to systems of action, where every employee, every machine, and every decision drives outcome.
We are delivering this through our manufacturing platform, which is based on three pillars. The first is Redzone, which is our connectivity and workforce solution. It is really empowering the front line by bringing data and information directly to workers on iPads or Android devices. This cuts out the need for workers to talk to supervisors and eliminates delays at the point where data and intelligence are delivered. The second part is Adaptive, which is our ERP focused on manufacturing. It serves as the intelligent backbone. The third part, which I am excited about, is Champion AI, which is our agentic layer. We look at it as amplifying human potential. These three form our manufacturing solution.
Regarding how we are delivering agentic AI, we think about it in three buckets. The first bucket is what we call persona-based agents. We have mapped out all the personas in manufacturing, from shop floor workers to supervisors to plant managers to inventory planners. Each of these people has mundane tasks today in the ERP or other systems. Nobody wakes up in the morning excited to work with SAP or their ERP. We have taken those mundane tasks away so agents do them, and these people can focus on high-value work. It is not just about high value; it is about how you increase engagement and get people excited about working with systems.
The second bucket is optimization. We have chosen very difficult problems to solve. One is inventory carrying costs, which is the third highest cost in mid-market manufacturing. This agent looks across the entire end-to-end supply chain and figures out ways to adjust replenishment levels, allowing the human to make the decision but providing very precise recommendations. The second optimization agent is our procurement agent, which cuts down a buyer's time by 50 percent. It is a dramatic reduction in wasted time but a tremendous increase in output. The third bucket is implementation agents. Gone are the days where systems should take 12 to 24 months to implement. We have developed an implementation methodology with agents that deploys mid-market, mid-size plants in 90 days. It is simple and very clear, and agents help with implementation through data migration, custom code, and other tasks.
In terms of architecture, everything is on AWS Bedrock for models and SageMaker for domain-specific training. We have Mongo, DynamoDB, and S3 for memory, and IAM and CAM for security and control. Regarding how we embrace flexibility as we thought about building agents into our offering, I have to say people are really excited. I do not think there is that much of a change management issue. The thing is, we have shifted from thinking about workflows to outcomes. The idea is not about scripting it all out; it is about setting the goals and guardrails and then letting the agent do the work.
Three things help with this approach. One is accepting probabilistic answers rather than deterministic answers. The second is creating multi-step autonomy rather than single-step autonomy. The third thing is putting a human in the loop. We are firm believers in that, especially for very critical decisions. You would not change replenishment levels without somebody really checking that. When people see what agents can do, like our procurement agent, when you run that and it addresses the messy inbox, they get excited about the possibilities.
Agentic Design Principles: Security, Governance, and Regulatory Compliance
In fact, this shows that a buyer saves 50% time. I mean, everyone wants to save 50% time. That's our experience. Fantastic, Sanjay, and I know it's great to hear your experience, Zach, at Stedi. Health insurance workflows are a little bit different. So how have you approached agentic design?
In a regulated industry, there are a couple of table stakes things that you have to respect from a security and compliance standpoint. You have to do things to a very high degree of quality. From an outcome standpoint, you don't want someone operating, doing heart surgery on day one. You want them to start with something very simple. For us, we looked at this in a couple of different ways. One is simplifying the technical complexity of what we were working on, which AWS helps us greatly with. The second is simplifying the product complexity.
If you can't sacrifice quality and security, you sacrifice scope. What that meant on day one was saying, look, what can we ship very quickly in single digit days to single digit weeks? And then what are customers going to be delighted if it works, but if it doesn't work, it's not going to have a bad outcome? You wouldn't want to put an agent that you're not quite sure about on appealing denied claims because maybe the claim is denied for a good reason and the agent is just going to take a lot of shots on goal to make that work.
So we put it towards eligibility checks. In that case, if somebody says their name is Will, and it turns out their name is not William, but actually Wilfred, there's not a really high cost in making that mistake. You think about the don't be evil versus can't be evil. We started off saying we're going to choose use cases where if things go wrong, they can't go that badly. The second piece was focusing on how we can avoid a lot of technical complexity by using the AWS tools that we get out of the box.
A lot of organizations look at these things thinking there's going to be a lot of special snowflake things involved in agentic development. We took the approach of saying, what is going to be the same? When you look at it, it's deployed with CDK and it's built with Code Build and the APIs are behind API Gateway and it's using IAM for identity and access management and Amazon Verified Permissions. So it kind of looks a whole lot like normal development. You just have to get comfortable with the pieces that are different.
Fantastic. And it's interesting, it's underpinning what can you take that you've learned, kind of lessons learned that can be applied, and then also around security and governance and how you think about that to move really quickly. So Zach, I'd actually love to kind of double click on how you have thought about security and governance in your agentic design.
Very similar to how we think about everything, we say first we want to make sure that it's starting from a principle of least privilege. I think that's something that gets talked about every year and everybody should have that top of mind, but just making sure that the agent doesn't have access to the vast majority of the tools that could be accessible to it. Stedi is an API first platform. I'll draw a distinction between that and an API only platform. We have many features that are available through our user interface, but an API first platform is a platform where anything that can be done through the user interface can be done via API or SDK.
AWS in that way is an API first platform. You can do a lot of things through the console, but hopefully you could do all those things via API as well. Starting from there, we use all of the controls that we have out of the box from AVP and IAM to say, OK, these are the ways that we control access to our APIs. We give the agent only the access that it needs, and there's great integrations available there. From there, we're layering in things like the Bedrock guardrails and making sure that there's appropriate guardrails on what the agent is and is not allowed to do and what sort of prompts that we're expecting from customers.
Fantastic, and I think Sanjay, you have a different kind of segment in manufacturing. I'd love to hear how QAD approaches security and compliance. All the things that Zach said are all relevant because our clients are also operating in highly regulated environments. The verticals that we focus on are automotive, food and beverage, life sciences. These are all areas where there are a lot of regulations that you have to follow. For us, security by design is very clear. When we designed Champion AI, it was basically that we design in such a way that autonomy doesn't increase while control decreases. So it's pretty clear that all the standard controls and checks from AWS are in place.
Those guardrails are there, including IAM full audit trail and KMS, ensuring that our customers' data is very important. You need to give customers the assurance that their data is not being used to train the model. That assurance is critical, and the big point is what Matt put up as well, saying "trust but verify." It's very clear to put the human in the loop, which is very important from a perspective of ensuring governance and control. It's fantastic to hear those underpinnings and those same best practices coming back again and again.
Measuring Success: Productivity Gains and Platform Expansion at Stedi and QAD
We're right at the end, so I'd love to close with a question for each of you. How are you measuring success of these agentic initiatives today? And if you think about the evolution of agents, how do you see that evolution evolving in the future? Sanjay, we'll start with you.
In manufacturing, it's quite simple. We look at very clear metrics: OEE, throughput, downtime, and cost. Everything has to be tied into specific measures in manufacturing. I'll give you an example. RedZone is deployed in 1,700 plants globally. We have 650,000 users that use it on a daily basis to make decisions. We measure production runs, and there are 6 million runs that we measure. We have achieved 26% productivity improvement measured through agents being deployed. We have 81% more engagement on the shop floor, and we have 35% reduction in attrition.
This is measured literally and specifically with the production runs. On the ERP side, we have specific things like the inventory agent measured on the reduction of inventory carrying costs. We've been able to record up to 30% reduction in inventory carrying costs with clients. The procurement agent cuts down time by 50%. Either you're looking at how you're increasing output or you're reducing costs. It's very simple in manufacturing.
Fantastic. Top line and bottom line. Absolutely. So Zach, I'd love to hear how you're at Stedi thinking about success and then the evolution of agents.
On the success standpoint, we're pretty astounded at the results. We're looking at it on two dimensions: how often are customers using this, and almost a third of our customers are using it every week, which is a pretty shocking number. It's shocking to us because when we looked at the number of people who were performing follow-up troubleshooting checks, it was in the single digit percentage. So it was a very small number of people who were following the best practices. Almost a third of our customers are using it weekly, and when you look at the resolution rate, it's again almost a third of the eligibility checks are successful. What that means in practice is that for every 10 instances that you look at, we're avoiding 3 phone calls, or 3 people being denied care and having to pay out of pocket, plus all the headache that comes from trying to reconcile those things afterwards.
We started small, but we like to start small and then expand from there. We built the agent in a pluggable modular way so that it could easily be plugged into other parts of the platform. It's not an eligibility troubleshooting agent. It's an agent that has an eligibility troubleshooting tool available to it. Since then, we've launched a payer selection agent. You might wonder how hard it is to choose the right insurance company, but Blue Cross Blue Shield has something like 33 different entities that it represents. Within that, there's a lot of further complexity in terms of whether it's a federal employee, commercial, Medicare, or Medicare Advantage.
We expanded into payer selection, and you can see this will move further and further out into the platform to things like claims and rejections, denials, and various sorts of resubmissions. What we're most excited about is what our customers build with us. We have hundreds of companies who build AI-driven revenue cycle management workflows on top of our platform, and many of them have asked whether, since Stedi is an API-driven platform, they can use the agent via API. There's an exploration there for us to figure out what exactly invoking an agent from another agent looks like. These are hot topics today.
I really appreciate the time today and just sharing your insights with the audience. We'll hand it back to Sean.
Four Pillars of Agentic Value: Speed, Adaptability, Quality, and Scalable Memory
Thank you so much, Zach, Sanjay, and Oluwale. That was a lot of incredible outcomes from Agente. Now that you have an agent-ready architecture in place, it's time to talk about taking advantage of it. When we talk about agentic value, we're not talking about marginal efficiency. We're no longer limited to those single prompt and response workflows I mentioned earlier that drive such incremental improvements. We're talking about orders of magnitude of value: faster decisions, shorter cycles, higher quality, and dramatically better utilization of both human and machine effort. Agents are what make this possible.
First, agents accelerate execution. In traditional workflows, every step depends on handoffs—a signal from one system, one team, or often one person to another. This creates lots of idle time everywhere and room for error. Agents collapse these gaps. They plan multiple steps ahead, pre-fetch context, and trigger downstream actions as soon as conditions are met. They're inherently asynchronous, coordinating and executing work in parallel instead of serially. This translates into much shorter cycle times, higher throughput, and far greater responsiveness to change. This is why they can pull multiple insurance providers at the same time, like you heard from Steady, or shrink ERP migrations from years to weeks, like you heard from QAD.
Second, agentic systems bring adaptive orchestration and decisioning. Traditional automation works from predefined rules or workflows: if X, then do Y. Agents work from intent. They can branch and explore alternative paths, test hypotheses, and converge on the best plan, all while staying within the boundaries you set. You're not hard-coding decisions anymore; you're teaching your architecture how to reason about those decisions dynamically.
Third, agentic architectures embed quality and resilience directly into the runtime. In the traditional world, quality assurance happens after the fact. Testing and validation is serial—it comes once the change has been made, pretty late. In an agentic system, these checks happen as the work is performed. Agents can validate those intermediate results, self-correct, and document provenance automatically. They can retry failed steps intelligently, simulate alternative plans, and even escalate if confidence falls below an expected threshold.
Finally, agents have scalable reach and persistent memory. They can pull context across those painful data silos you have in a business, across formats, across languages even. They can reason over structured data, text, images, or telemetry and unify it all into a single decision flow. More importantly, they remember. They retain what worked, what failed, and what improved the outcome. Over time, that memory that agents can have compounds. It allows the system to get smarter—not because the underlying model is changing, but because the system as a whole has learned from its own operations. That's how you see compounding returns, the kind of learning loops that continuously drive down costs while driving up quality and speed.
We opened up today with a look at the art of the possible—that fun orchestra picture with thousands of agents working in concert to complete complex tasks. How are we going to get there? Baby steps. Today, we're moving from an LLM-assisted workflow that maybe gives you five to fifteen percent efficiency gains to agentic workflows like this one that accelerate outcomes and lower your costs. It starts with incorporating agents in discrete parts of a business process to augment and improve work. Over time, it expands to more agents handling different parts of the process. Eventually, we expect processes that are fully agentic, able to deliver extraordinary results, much lower costs, and much faster cycle times. It sounds unreasonable, but it's not. I'm already seeing it today from our bleeding-edge customers. If you attended the Amazon innovation talk earlier, you heard that we've deployed more than twenty thousand agents across our own business since July to automate a huge variety of workflows.
This is where the industry is going. Now, one way that every business can benefit from agents right now is with engineering development. Dave Treadwell this morning talked about the speed AI native development is unlocking for Amazon, where agents are fundamentally transforming the way we build. In an AI native model, agents don't just generate code, they orchestrate the entire life cycle. Bringing product people, security people, operations and engineers together in real time.
Instead of silo teams discovering issues as development progresses, agents surface those trade-offs early, mediate decisions, and keep everyone aligned, so the system evolves holistically from the start. That's the real breakthrough. Agents turn software development from a series of disconnected handoffs into a continuously collaborative, intelligence-driven process. Customers adopting AI native development processes are achieving extraordinary results.
IT services company Repro developed a production-ready enterprise healthcare platform in just 20 hours' time, leveraging domain-driven design. And trading platform Don originally scoped 2 months to build new capabilities, but by adopting our AI native practices and developing, they developed those capabilities in 48 hours. That's truly incredible. Beyond development, we see a lot of powerful agentic use cases specific to every industry.
Pattern's E-Commerce Transformation: AI-Generated Product Photography and Conversion Optimization
Whether it's telecom giant Ericsson using Bedrock Agent Core to drive double digit productivity gains in R&D or Honda using Agent Core to speed production of its autonomous driving technology. That's in addition to what you just heard from Steady and QAD. Now, another great industry example is our next speaker from e-commerce company Pattern. Pattern has reimagined business processes with agents at the core to drive significant results. To share more, please welcome the CEO of Pattern, David Wright, and our leader of technology in Europe, Melanie McGrory.
So Dave, welcome. Thank you very much for being here. You started in 2013 and you've come a long way since then. You're now the number one seller on Amazon Marketplace. Tell us a little bit about Pattern and a bit of background about how you got there. Yeah, it's a fun story actually. I had a cousin, and she had created this little girl's headband brand. Sitting there going through spreadsheets, my background is all in the tech and data science side. So I'd never sold a widget, never done any marketing, and matter of fact, most of those things sort of bothered me.
She started talking about marketing and some terms I'd never heard of like ROAS. And it was intriguing from a data perspective, so one night going through spreadsheets, next thing you know, I was helping a few other brands, and now we have Pattern, and we just finally made it through the going public process. I wouldn't recommend it to anybody. Like Amazon, we spoke earlier about the fact that you work a bit backwards from the outcome that you're trying to generate. Can you tell us a little bit about one of the problems that you're trying to solve for?
Yeah, it's been a really fun journey. I'll give you one example, as you can see on the slide here. I was meeting with a brand called You Theory, you may have heard of it, it's owned by Jamieson up in Canada. And I was meeting with their board because they're spending about 19% of total revenue on ads. And they're growing nicely, and I tried to tell them I was like, you have a core problem. When people go to your products, they convert 10.4% of the time, but your competitive set of products, they convert 16% to 16.7% of the time.
We were stripping out all branded comparisons. This is a very like-for-like number. And I was like, so your growth is severely limited by this fact. So of course, it's a very complex problem to try to figure out: OK, what are we doing wrong? Is the You Theory branding itself flawed? Is the product flawed? Is the messaging flawed?
What could we do? Because it's so complex, and then you start thinking globally, this is not just a brand that wants to sell in the US and they don't just want to sell on Amazon, they want to sell globally. So this became a problem that we solved leveraging agents. What we have built, and I didn't imagine in 2013 that this would be possible. We have 29 patents either issued or pending on the tech stack itself.
What we do is we set up sensors anytime you have a data-rich environment where you have the ability to create a sensor like this , so a negative 6 percent conversion to competitor would be a sensor. Once I've triggered a sensor, I need to go and collect information. We have an information collection agent that will go and interrogate the web and interrogate all the sources, pull in data and say, this is what I'm working with. Now we ask, what are all my competitors doing? That would be your shelf, because obviously they're beating me by 6 points. So now let's analyze what they're doing with information collection again, put this together into a strategy.
Then we start saying, we need to be faster at generating product photography. In this case, once you've interrogated the web and you've pulled the images in, we need to be able to do AI-generated product photography. But if the image quality wasn't good enough, we actually had to build a piece of hardware. You can see our robotic arm, which we call the Portal. It's on the left there. We will actually shoot product photography, and you can see our turnstile table with a motor in there. We've had a lot of fun with this. Those are LED screens where we'll actually train a LoRA model on a lot of different shots from the same angles. We'll do hundreds of shots inside that Portal. Then we're able to do AI-generated product photography, and you can use an agent to audit it.
So then it can go back through and you can say, here's a brand style guide, here's some things that you can and can't say, here's the actual dosage from their PIM. Then we can air check it using an agent. All of these things, and then if we get some part of that image wrong, we could regenerate not the whole thing, we could regenerate parts of it using a content creation agent. This has resulted in some great results for You Theory.
With the agent's view, that's been a game changer in terms of what you can deliver to your customers. You know, Shao mentioned that accelerated outcome, and this is really a great example of that. If we then look at the larger picture, we have a technical audience here and I'm sure we're all itching to know about the architecture that you use in order to be able to solve problems such as this. Can you tell us a little bit about that architecture?
Of course, I'm sure most of you are pretty familiar with it. I've got the slide up here with the high-level architecture, but a couple of points that I think are interesting. In the old world of a non-agentic workflow, you start with an input, you have an output, and you have steps 1, 2, 3, 4. In an agentic workflow, between steps 1 and 2, I can now make a decision on exactly where I go, leveraging a set of data that we might have. So you can see we have PXM there. I can query an internal information system that we might have and data and feed that into a model. Then I can let an agent make a decision as to whether I go on from 2 to 3 to 4, or I skip 3 altogether.
We also have long-term and short-term memory, so in that process, we can remember what was done before and what might be more impactful. I've been involved in tech my whole life, and I actually never thought we would be able to break this type of barrier in terms of just efficiency and speed. We've added a few things on here that are not Amazon and Amazon lettuce, so thank you for everyone. But I'm sure we're all using products that are Amazon and non-Amazon. The great thing about the core, the bedrock, the foundation that Amazon has built is that it does allow for the most part full interoperability with other systems.
In a sense, it makes it seamless to blend it all together and enables your team to be able to deliver what is needed. When you apply the architecture and you're thinking about the solutions that you're trying to generate, give us an example of maybe some of the results that you've seen from this.
When we put this slide together, the Amazon team came back and said, "So you only went from 15% conversion to 17%?" And I said, "That's a massive number." This is across the Pattern portfolio as a whole—100,000 SKUs, 60 marketplaces globally, all running through these agentic flows. A 2% increase in conversion is not just linear math. If you think about it, revenue is essentially traffic times conversion times price. You would think it would be linear, but if you can increase conversion, you actually increase traffic as well. So that becomes a massively impactful number.
And then, of course, there's the reduction in cost. We used to rent houses so that we could do photography of a bottle of supplements in a bathroom and then in a kitchen and everywhere else. Now we've completely changed the game for us with a 96% reduction in photoshoot time. Well, we would say you may be a little bit scrappy, but scrappy works because you are a huge success story. Can you give us an example of something that you've generated?
Well, the Youth Theory product is collagen. Its main use is for what people latch onto—really good hair and young skin. Yes, I think that's a little bit artistic license. I really like that one. When you're building workflows with agents at the center to drive the results that you want, what are a couple of things that you've learned that you would give us some examples for builders to take away?
What are one or two things that you would recommend to all of us here today? Well, maybe a couple of things. If this becomes a people game—and I think we talk a lot about the technology, but it's always people who are implementing it. I see companies out there saying, "OK, we're going to do a reduction in force because of AI." We're completely the opposite. We're increasing our technology budget faster than we ever have in our thirteen-year history, and we're somewhat obsessed about paying the highest prices.
I think that companies are somewhat delusional if they think they're going to come in at the average pay for an AI engineer based on a salary chart. If that's what you're paying, that's probably what you're getting. Every company I talk to says, "Oh, we hire the best." But statistically, are you really going to be hiring the best if you pay bottom quartile versus top quartile? I think that in the next two or three years, the companies that don't figure that piece out are going to get left vastly behind.
The speed of what we're talking about—96% reduction in photoshoots and whatnot—I think that will be key. And then there will be leaders you'll see that are historically probably great leaders. I had one executive who I thought was a phenomenal leader, but I just couldn't get him into the technology. It was almost like saying, "Hey, don't you love pizza?" and he says, "No, I want to do something else." This is now a game where if your leaders are not into and completely immersed in AI and the technology and the benefits, they will get left behind. You probably need to go in a different direction, or you'll lose. So I'll probably just say at the end, this just becomes a people game.
When we think about people, can you tell us a little bit about how you created that image? Yeah, we'd love to hear a little bit more about it. What did you do? Well, you saw the portal, so first you can see the bottle of collagen. Once we shot that in the portal with the robotic arm, it'll take hundreds of images from different angles. And then once we have that right, we'll audit that bottle of collagen. You can see that little molecule right there—they call that a molecule. That's very, very difficult to do. And if you run that through Gemini or OpenAI without being portal enhanced, sometimes there's just not enough good imagery out there, so you actually have to shoot it yourself.
We did that, and then we just layered in Shaun's hair without even enhancing it. What about how you've approached where to deploy agents? From your perspective, how do you know which use cases would deliver the most value?
I would say something I probably never thought I would say is that it's almost like a level of creativity. If you have a standard workflow where you know your input and output, that's one thing. But if you have a really good idea of what the output ought to be while the solution in getting there can be variable and creative, that is a perfect spot for an agentic workflow. You're very clear about the outcome, but the way you get there and the data that feeds the input can be creative. If data can influence that output, then an agentic workflow becomes the foundation for success in the future.
I want to thank you very much for being here on stage. We heard a lot from Shaun about how people are using agents today and how other companies are using them to accelerate business outcomes. With that, we're going to hand back to Shaun to take us home and wrap up a bit more about agentic AI. Thanks very much Dave for joining us on stage today.
The Future is Agentic: AWS's Comprehensive Platform for Building Trusted AI Agents
You bet, thanks for having me. Thanks Dave and Melanie, and for the streamed audience out there, especially my wife, check the bathroom Angela—there's no collagen there, that was AI generated if there's any confusion. We are at a major inflection point in technology and the agentic advantage is real. Agents are turbocharging operational agility and unlocking extraordinary business value. To realize your own agentic advantage, AWS is the best place to build and deploy trusted and performing agents anywhere in the world. I told you about everything we can do today with Bedrock AgentCore to run them in production safely, securely, and scalably. We saw tons of new capabilities announced today, including evaluations and policy. Even beyond AgentCore, we have so much for you—everything you need.
Whether it's ready-to-use agentic applications like Kiro for software development, our new security and DevOps agents, AWS Transform to accelerate migrations, Amazon Connect to give your customers the best experience possible, or Amazon Q to be the best teammate for your employees. We support the most popular frameworks for building, including open source capabilities like our own Strands Agents. We have the broadest set of model choices in Bedrock, including all the new models you heard Matt announce this morning and our own family of the new Amazon Nova models, and the very best ML infrastructure with SageMaker, including our new Nova Forge for customers to build their own frontier AI models through an open training approach. And not least, we have our purpose-built silicon for AI training and inference, including not just the latest NVIDIA-powered P6s, but also our new Trainium 3 Ultra servers. Do not miss Swami and Dave's keynotes this week to hear what else is coming.
Now, it's been ten years since my first re:Invent. Back then, I couldn't imagine the changes that cloud would bring to business. I saw firsthand how enterprises were first disrupted and then accelerated by the power of building in the cloud. We're there again. With agents, we're starting to realize that agents are going to be engaged in every business process, every workflow, and every system. They are becoming foundational and it's going to change how we all work for the better. It might sound unbelievable, but if history is a lesson, it's going to happen.
With that, I want to thank our incredible customer speakers, my colleagues Melanie and Olawale, and I want to thank all of you for taking the time to join us. Please go check out more great sessions, have a great time learning, and have a great re:Invent. Thank you.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)