Satya Nadella Keynote: Microsoft AI Tour Sydney

Right. Good morning. Good morning. It's fantastic to be back in Sydney and in Australia. This morning, I got up, and then I tried to sort of go down memory lane, thanks to Copilot, and try and get back to the first time I showed up in Sydney for one of the developer conferences. And it was, I think, in 1993, and I remember distinctly getting through the airport with my luggable Compaq computers, and the conference was all about -- we had introduced, I think, for the first time back then, Visual Basic for Applications. That was like we replaced the macros in Excel 4. Excel 5 had come out, and we had Visual Basic. And I distinctly remember, in fact, having a session. This time around, there a few more people for my session, and then I remember going to one of the banks and having a big discussion on going from macros to Visual Basic -- is it a good thing, or what have you? But little did I know that whatever 30 odd, 35 years nearly after that, the hardest thing in tech is, again, guess what? Agents in Excel and or command line interfaces for coding agents. I mean, it's fascinating that as the frontier moves, so does the arrival back to where you started, to sort of use that T.S. Eliot metaphor.

To me, I want to spend most of my conversation today and time today to talk a little bit about that new frontier, and how are we all collectively going to push that forward with essentially what is a new input? The outcomes that matter still remain very much the same, right? When we think about any organization, any institution, even individually, what we care about is the excellence in outcomes, right, which is how do we show up and change that engagement with customers, or the employee experience, or the efficiency with which any given business process gets completed, or even bending the curve in innovation, right? Whether it's software development or drug discovery, how are we going to truly change that? And that's sort of the frontier that we are trying to move. And if you look back at it, right, throughout these transitions that we've had in technology, a lot of this has come down to us changing the way we work with technology, but more than the way we work with technology, how technology reshapes the work we do. That's sort of been the story of sort of the movement on the frontier, and it starts, quite frankly, with that mindset, right, that curiosity, the posture that we need to have that when there's something new, you go all in; you learn it. That means you really need to think about your skills. You need the new tools. And perhaps one of the most interesting things and important things that we'll talk about in the AI era is codifying the outcome, perhaps even more robustly going forward into these evals, right? If you think about AI, anyone who's sort of doing applied AI starts by sort of with this mantra called the product is the eval and that hill-climbing on evals, right? And that's that subtle, nuanced way all of us judge what great looks like, good looks like inside of your organization for any outcome becomes the real currency through which we can then change the use of technology and have it reshape our organizational frontier.

And so to that end, we are building this end-to-end tech stack, right? It's got three parts to it, and for the rest of my talk, I want to sort of unpack each of these in some detail, and the true line across all of this is going to be two things: intelligence and trust. Because ultimately, you really want the intelligence to show through across each layer of the stack and amplify what you're doing, move that frontier, move that outcome, but you really need trust because it quickly can become vicious cycle versus a virtuous cycle if you really don't have intelligence and trust. And so we'll sort of talk through how those all come together. So let's start with sort of the top of the stack, which is around these high-value scenarios and high-value applications or agentic systems. And for us, it starts with the way we are architecting Copilot. If you think about Copilot, whatever, three years ago or so, we started with mostly a chat-based response, request-response type of pattern, and especially with GPT-4, it became a new way for us to discover knowledge inside the organization or on the web. And then subsequently, we had reasoning models come out that sort of helped us not just have essentially new search, but we had reasoning plus search, right? Even the chat interface started to pretty drastically change with reasoning models. Then we started getting better at sort of long-run agents that you could send tool-calling. And so that started sort of really the beginning of agents like Researcher and Deep Research and what have you. So the first thing we started doing was to say, okay, let's take even chat and that reasoning and make sure that you have access to all the models, right? So in fact, when I use Copilot today, I'm mostly sort of doing Auto, and Auto is picking the best model for my intent, and then I have a dropdown with all the models available. And that is sort of how the chat interface itself has evolved, right? You used to be, in the beginning, you used to have a router, but now the router is becoming smarter and smarter in being able to sort of do the intent classification, pick the right models for the intents.

Next, we've now taken that idea of having multiple models and put them even in the context of agents. So if you go to Researcher, you now have these new modalities. For example, you have a modality called Critique, and I love that a lot because, after all, how do we work with people? We critique each other and our ideas. That's the way you work with even models. In fact, another mode that I like a lot is Counsel, which is essentially assigning a particular task or a research topic to a group, and then you get the best of all the models come together. It's fantastic to see it if you have not used it to sort of look at both the models, not only doing critique, but also finding biases in each model, and then, in fact, insights from each model, right? So when even the research report comes back, you gain deeper insights because that you have multiple models at your disposal. Now, the thing that I'm really excited about is this next big thing that we've done, and in fact, today's the general availability of agent mode across Word, Excel, and PowerPoint.

Yeah, it's kind of like 1993 again. And it's actually, in some sense, all of my life's work is right here. And the reason I say that is because if you think about what's happened is Excel became a super popular tool, because it helped us intuitively get number sense, right? It had representational power in the canvas, right? The rows and columns have information content, right? As soon as you see a spreadsheet, you kind of make sense of what that spreadsheet is, right? That's why the canvas has lasted this many decades, in fact. And so now think about what agent mode does. What agent mode does is it takes that information in the sheet, the positional or spatial information representation, and then mixes it with essentially these powerful reasoning models and the ability to take action -- not just reason, but take action. And so when you put that together, it's magic, right, which is you now are creating a very -- you're instructing the model to create a very sophisticated model, an Excel model. And then when it creates the model, it's represented in this canvas, which you know how to even spatially inspect, and you know what's driving what. What's next to what? And then you can then further reason about it. And in fact, you can take that entire reasoning trace, plus the sheet, plus the model, and share it, right? I mean, think about the cognitive work inside the company and how it's getting reshaped. It has familiarity of a spreadsheet, but everything around it is changed. And that, I think, is when you know you have a powerful sort of set of technology. And so we're really, really excited about this agent mode. And of course, we're not stopping there, right? The next thing that is also built into Copilot now is Cowork, the ability to now assign a synchronous task in a queue to Cowork, which, by the way, could be creating a Word document or an Excel spreadsheet or what have you, and that then you can open up in agent mode and round trip this entire way of working. It does feel like when we move from faxes and interoffice memos to emails and spreadsheets and documents floating around the enterprise. It feels like that time again where the way we work, the artifacts we create, the workflows that we are involved in, are fundamentally going through a sea change. And to kind of give you a little bit of a flavor for this Copilot ecosystem at work, I wanted to introduce my colleague, Nirja. Nirja, take it over.

Thank you so much, Nirja. It's so wonderful to see even the evolution of Copilot in just the last few years, right? I mean, in the first instance, it was a bit of a UI on a model, but now you can think of even the Copilot system as essentially a cognitive loop that or a harness that you have at your disposal for you to be able to have more agency, more ambition on what you want to create, how you want to, in some sense, initiate the trajectory of a task. And that, I think, is what you see in all of those features, and then it's also becoming quickly multiplayer, right? So you're not just single-player mode, but when you're in Teams, you're interacting with not just Copilot, but all the other agents as well. And that, I think, is what is going to probably unfold very rapidly throughout this calendar year. Now, we are seeing fantastic momentum. The great news is this is not about sort of this rate of diffusion here happening slowly. It's in some sense happening all very fast, everywhere, at the same time. In Australia, we've now got some fantastic momentum around Copilot usage across all of the top companies that are represented on this slide. I think 17 of the 20 top companies in the exchange here are using Copilot deeply. All the banks are using it. Multiple. I think there is 57 government agencies using it. So it's fantastic to see the momentum, and I had a chance yesterday to meet with this team at Australia Post that is sort of looking at Copilot and the adoption of Copilot, and it's a complete bottoms-up way. They have all these Copilot champions and team captains who are really making sure that everyone is trained on it and is getting the best out of it. But the thing that stood out for me is how much care they're sort of taking in making accessibility one of the big beneficiaries of Copilot. In other words, they've used Copilot and its native technology, but it's really driving everyone with all the abilities to be able to fully participate and give them more agency inside of the enterprise, inside of a Teams conversation, in a meeting, inside of being able to sort of use Copilot to really get things done in a much more seamless way, and it was fantastic to see that. So that was the first part of the stack. So we go to the next stage, which is the AI platform or the agent platform. Now, the best way to think about this is in a couple of different layers. There is models. There's context. There is agents. There's harness, and then there's security around it, right? So you want to have access to all the models. You want to be able to have access to all of the rich context. You then want the agents to be built on top of that, right, using any model with all the context, and then you want the harness that knows how to orchestrate all of it, right? That's sort of the core of it, and then, of course, you need the security and compliance around it. So I want to unpack a little bit of this platform. And for us, the organizing sort of piece here is Foundry, and Foundry brings all of these things together, whether it's the agent and the multiagent run times, whether it's models, whether it's the context layer. And so I kind of want to walk through each one of these pieces in some detail. The first is the models. Of course, you have the frontier models, but the reality is, we have 11,000 models in Foundry today, and this is sort of the base material for you to be able to build your own agent in agentic systems. In fact, all the frontier models, as I said, are there, and they're rapidly evolving. We ourselves, with MAI have launched, even in the last, I think, month and a half, a fantastic transcription model, an image model in two varieties. One is a very high-performance and low-cost image model, a new voice model, again. And so it's fantastic to see the, really, not just the big frontier models, but even the helper models we use all the time for different modalities inside of our applications are all rapidly evolving. But I wanted to talk about this one model that struck a chord for me because what is the innovation agenda around models, and how does one go about it and build this as a capability that truly empowers the next level of innovation? So to showcase that, let me just roll the video on GigaTIME.

What was said in the video sort of makes a ton of sense, that you can take a technology like this, take what is a complex test that costs a lot of money, takes a lot of time, and truly democratize it, right, that point about every tertiary hospital in every secondary city now with just a pathology, a regular slide pathology can get this test done, essentially, is a pretty big breakthrough. But the question is how that came about. The way it came about is two research institutions that have had this knowledge, right, this ability. They've obviously had all this data, and they were able to take that data, train a model, a generative model, to essentially do simulation of a test using compute and share it with the world, right? That, to me, is what's happening, right? This is not about one frontier model company doing things. It's about all of us as an ecosystem at the frontier have the ability to use these new techniques, like generative AI, these new algorithms, this new compute infrastructure that's there, and turn what is essentially the tacit knowledge we have, the information we have, the data edge we have, and turn that not only to move the frontier inside the organization, but you can move the frontier for the society at large. And that, I think, is what's exciting about sort of the true democratic force that's underlying all this. The next thing that I want to talk about beyond the models is context. In some sense context around the models is perhaps one of the core assets that's, quite frankly, distributed, right? Every organization has unique context, and you want to be able to organize that context such that you can put it together with these models to both reason and take action. And that's sort of one of the top engineering considerations and architectural considerations, Work IQ, right? So Work IQ is that rich database underneath Microsoft 365, right? Everything Nirja just showed around Copilot was tapping into Work IQ, and that Work IQ is available headless through an MCP interface for any app developer to be able to build an agentic system around it, and this -- think about this. Up to now, I've always said that one of the most important databases inside of any organization is the database we don't talk about, right, which is the database that contains all of the people and their relationship to other people inside the organizations, the communications graph and data, the document libraries, project files -- all that is in Work IQ. Similarly, another very important database when it comes to thinking about outcomes is what's in Fabric, right? If you think about all the various dashboards that have gotten created over the decade, and if you think of any dashboard is some outcome, some metric, someone's measuring. So there's a rich semantic model that was created. That's, again, super important context to be able to give these LLMs and models as you try to drive your workflows. And by the way, Fabric now has the ability to reach out to any data, not just in Azure, but everywhere, right? You can connect to AWS. You can connect to GCP, bring all of the enterprise data asset as context inside of this Fabric IQ. And then, in fact, the outer loop of this can all be Foundry IQ. What Foundry IQ does is essentially even not only gets you that document database. So if you have a lot of unstructured information, you can then vectorize it, make it a part of your context. But you can also use it as an outer planner to be able to plan across Work IQ, Fabric, and Foundry, right? That's sort of how these things come together. So context and this IQ layer beyond the models becomes the other second important piece. Now the third thing I want to talk about when it comes to agent platform is the agent service itself, right? With every shift in platforms, you kind of need a runtime, and that's what this is. The new runtime for AI agents is this Foundry agent service, and really exciting announcements today up and down the stack, right? We have now persistent memory. We have tools use. So basically models that then can be attached to knowledge that can then also have attached to tools. You can then, on top of it, build these multiagent and agentic systems that you can orchestrate using any orchestrator you choose. You can publish them with one click. And then, as I said, you have memory that's persistent across all of this, and they're long-living agents, and they're stateful agents. And so the one thing that I'm most excited about is the hosted agent service that now goes into preview mode as well. So it's really exciting to see this because now what do you have? Okay, all right. In fact, this is like that day we started the cloud journey the first time we talked about containers and said, oh, there's a new type of virtual machine, effectively, for cloud-native applications. So think of this as the birth of this new generation of AI-native applications with a hosted service. So literally, you have a sandbox per session that can bootstrap itself fast, right? So it's instant on. So in fact, when you look at the service, you will instantly be able to bring up a sandbox. The sandboxes are isolated. You have persistent memory that is session state is persistent. And then, of course, in like any other service in the cloud, you can scale up and down elastically, right? So this becomes, I think, a massive boost because one of the things I've been talking to a lot of developers, and they come and say, look, when can we have a hosted service for agents? Because agents are new software, right? People talk about what's happening to software. Guess what? All software is going agentic, and you need to host these agents. You need to have them have their own computer, essentially, and that's what this hosted container sandbox is. And then it needs to have state, so that's kind of where the persistence comes. You need to be able to scale them up and down. And so we are very, very excited about this new Foundry service with hosted agents and what you all can do to scale the agentic systems going forward. Now then you have the runtimes, you have the models, you have the context -- you want to start building these applications, and that's where we have the richness of multiple different tools. You can start right inside of Copilot with something like App Builder, hit a prompt, and here's a career path app that you just built with a single prompt inside of a chat session. It's a full stack app that gets generated. You can share it, or you can go to Copilot Studio, again, build an agent using a simple set of prompts. In this case, it's an onboarding app that got built as our onboarding agent that got built using Copilot studio. And then, of course, you have the full richness of the professional developer tools, right, with VS Code and GitHub and now all the agentic tools. In fact, one of the most exciting things is right here in Australia, we now have, in GitHub, 2.3 million strong community. Yeah, it's and it's growing, right? It's 26% growth year over year. It's great to see the community continuing to scale. And GitHub is not just about the code repo, but it's also now becoming the place, the control plane for all the agents. So we described it as Agent HQ, so you have every coding agent that that you use will be available to you right next to your repo, right? So you have your code. You have your coding agents. And in fact, one of my favorite new features is, not only do I have my source code, but I have my session traces as a first-class thing inside of GitHub, right? So because, after all, going forward, it becomes very important for us to not only have the code, but also the traces that generated the code because of the coding agents right with the code. And I think this become, in fact, for true provenance of code going forward, that becomes a pretty important attribute. And then we're also exposing the tooling through all the form factors, right? As I said, CLI is back. Everybody loves it. And the reason why is not just because a command line is powerful, but it's command line when coupled with rich agentic models and natural language that have access to everything. In fact, one of the coolest things that I found myself doing is you can literally go to your repo in GitHub CLI and then you can connect to Work IQ through something like an MCP and say, hey, I was in a design meeting yesterday. We had a discussion on this topic. Can you make sure that my repo is in sync with the discussion? I mean, think about that. How crazy is it, right? With a natural language utterance, essentially, you're able to relate a design meeting that was mostly a transcription -- that to source code, all being done in a command line, right? That's the type of richness in that form factor that you can bring. And then, of course, IDEs still matter. In fact, IDEs are becoming ADEs, right? So they're now evolving to, basically, now the sessions app inside of VS Code is basically an ADE. And sometimes when you go into an agent mode and or the agentic output, it says please open up in VS Code. And if you want to inspect any code, you want that profiler there. You want the debugger there. And so we have all the form factors that developers need and want and want to use as part of a workflow all available to them. And so to show you all of this developer richness, I wanted to introduce Australia's own Damien to the stage. Damian, take it away.

I must say we have come a long way in our naming as well from Visual Basic to rubber duck agents. I love it. It's pushing the frontier. And one of the other super important considerations, right, when you build all these agents, right, what Damian showed is just pretty stunning, right, between Copilot Studio plus what he showed, you can be assured of one thing, which is we're going to have lots and lots of agents inside of our organization, in our tenant, in our network. And so one of the key considerations is going to be security and compliance and observability, and that's what's the motivation for Agent 365, right, which is how do we build a new control plane, just like we did for the previous era, but for the agentic era, right? We need an identity. So we extended Entra. We needed security. We extended Defender. We needed data security. We extended Purview. And so fundamentally, now you have in Agent 365 a platform and a control plane. And by the way, they don't need to be hosted on Foundry. They can be hosted on Azure. They can be anywhere, again, on AWS, GCP, anywhere on your private cloud, even, but you have this one control plane that you can use to be able to manage all of these agents and have real visibility into all of these agents, and set governance policies and enforce those governance policies. Now, one other consideration that perhaps in the months to come and definitely in the years to come will probably become the top consideration is in this age of AI where you no longer just have a platform, but you have a platform that extends into a learning system. So then the question for any software company, any new AI startup, or, for that matter, any bank, insurance, energy, telco, whatever -- pick your favorite sort of sector of the economy -- is, guess what? There's no limits to these learning systems. They can pick up what's novel and new and interesting to know about any business and just make it part of their base weights. And so this is a pretty new thing that's never happened before because it's not like the last time you thought about a digital system. That digital system was a self-improving learning system. It was a system that you could sort of use with your context and your workflows and make sure that that IP that you create on top of it is yours. And that is what you now and all of us have to do, even in the AI era. And to do that, that means the first thing is to really think about each of these layers of the agent platform and make sure that you start with your evals, right, after all, your hill-climbing with any agent system you build on some eval and making sure that those evals are private. In fact, perhaps one of the most important IP of an organization or a software company that is doing something unique is going to be captured in some private benchmark, right? In fact, all the public benchmarks are so saturated they're more or less becoming sort of noise. And so therefore the real game for anything is about hill-climbing on benchmarks, and private benchmarks you create, private evaluation sets you create, are going to become the new IP. The next thing you want to make sure is you decouple, right? You want to have access to multiple models. Like rubber duck is cool, or critique is cool, but you want the harness, the context, and the models to all be decoupled. It's tempting to couple them to get some gain on an eval, but then you're stuck with that one particular local maxima. But you're not going for the global maxima, where you want to be able to ride every frontier gain or your own model development. You want that context that's there with your -- you want to externalize as much of the expertise inside of your organization into skills, right? Simple markdown files like in Cowork, one of the cool things is you start writing these skills, and you even compound the skills individually and at an organizational level. So it's kind of externalizing the expertise of an organization. You want to keep that, again, very private. You want to take all the tacit knowledge of the organization. In fact, the tacit knowledge of the organization is with the people and the way they work and the way they work with AI. Those traces when I talked about in GitHub, now in sessions, you have the traces and the trajectories of all of your agents. That's super important data. Not just your source code, but that data is your data. And you can train your own model. You can take a base model, which is a pretrained model an open-weight model, and train that with these traces, and then put it back in Copilot in Auto, and it'll then route to your model. So I think that this idea and this consideration around sovereignty of a firm because, ultimately, what is the future of a firm? The future of a firm is its ability to continuously, monotonically get better at its knowledge creation, and it's able to capture its tacit knowledge that is not only inside of context or in skills, but even in the weights that you control as your own IP. And so that's what we're building the agent platform for, and you'll continue to see us, even going into our developer conference in early part of June, talk about this in great detail because we think that this is what, as I said, any startup, any software company, and, for that matter, any part of the economy that now needs to be thinking of all of these pretty strategically. And sort of speaking of that in Australia, we have fantastic momentum in terms of people building these agentic systems. I had a chance yesterday to meet with many of them. AEMO has taken Fabric and Foundry and built out a fantastic system to be able to improve the signal-to-noise ratio around alerts, right? Because how do you really have AI help them manage, in fact, some of the most critical infrastructure around the grid in a much more streamlined way and an effective way? Had a chance to sort of meet the team at the Commonwealth Bank, where they have built and extended both their customer-facing bot with all the new power of these models, but also for their agents in their call centers to be able to take the escalations and have the tools. The productivity gains that they're seeing is tremendous. I had a chance to meet with even one student and as well as a software company called Cogniti that's working with the University of Sydney in enabling, in fact, teachers, right? We know that a lot of the students are using these tools to learn, but teachers being able to create new course material and change the way they conduct their sessions and so on were fantastic to see. And of course, I had a chance to meet with Cricket Australia. And look, I mean, it's always awesome. I think Australia has been collecting -- and Cricket Australia -- has been collecting stats from, I forget, 1896 or something like that. In fact, the last time I was here, somebody was telling me, hey, if you ever wanted to check the scorecards of Steve Smith when he was a 15-year-old, there's no problem. You have that all data. And I'm a big cricket stats buff. And one of the things, in fact, one of the queries I have, which I'm going to do this weekend, is -- and in fact, I was asking them whether they have basically an MCP server for all of their data. And it's all in Azure Blob Storage, so I must assume it is there. But one of my things I want to check out is the conversion ratios. If truly Steve Smith's under-15 scores are there, I want to see them next Vaibhav Suryavanshi's scores and see how good is this guy going to be when he grows up? And so that type of ability to use data in a rich way and then translate it in their case to the outcome that they desire, which is fan engagement, is fantastic to see. So let's roll the video.