Developer Keynote | Dreamforce 2025

Please welcome SVP Trailblazer Relations Salesforce Christoph Konrad.

All right.

Hello everyone. Welcome to the developer keynote. It is an incredible time to be developers. But before we dive in, I want to take a moment to thank you. Thank you to all of you for being here today. A lot of people in the room and online, but most importantly, thank you for being the best developer community in the industry. I see MVPs here, I see golden hoodies, I see developer user group leaders. Did you know that we now have 140 developer user groups in over 90 countries? And, and this is just one of the many, many things that you do to come together, learn together and support each other. Thank you for everything that you do. And yes, please give yourself a big round of applause. Now, together you are also driving incredible impact and innovation. I know you have seen a version of this slide before, but this is really the developer version. And the point is, this is your impact. You are building all this and the impact is massive. Thank you for everything that you built on the platform. Now, we are here today because our developer world is changing. In fact, there are three shifts that we are experiencing today that I would like to discuss with you in this keynote. The first one is that agentic AI is really transforming how we build. And you saw it in the main keynote yesterday. We are moving from you writing every single line of code to a world where you are using natural language to describe the outcome that you want. This is intent driven development. If you want, this is Vibe coding. And if you have tried it, you know that it feels extremely empowering. It's so easy to create these prototypes in no time and impress your friends, which is important, right? But there is a problem, there is a challenge, and that is how do you Vibe code beyond these prototypes? Because in the enterprise, let's face it, we can't just let AI generate random code. We need security, we need trust, we need governance. And if you want that code to be really useful, it has to be grounded in your rules, in your own metadata. That is exactly why we are so excited to introduce AgentForce vibes. This is vive coding for the enterprise. It is everything you like about Vibe coding, plus that enterprise grade security that's built into the platform and it is grounded in the metadata that you built over the years in your orgs, which means that the result is, is code that's actually useful, that is secure and that you feel comfortable deploying. But you know, agentic AI is just not changing how you build, it's also changing what you build. We are Moving from building apps to building agents. Now, apps, they are deterministic. They do exactly what you coded, at least most of the time. But the agents, it's different because they reason, they orchestrate the act by themselves, which is awesome. But it also means that they are non deterministic and sometimes not that predictable. That is our second challenge. How do we make these agents more predictable? How do we bring control to something that is just reasoning by itself? And that is why we are so excited about AgentForce 360, because it makes your agents more predictable. First by grounding your agents in your data and your metadata. But now with hybrid reasoning, it really gives you more control over the level of determinism that you want. Now these agents, they are still going to perform tasks that you never anticipated. And that's what makes them so powerful, because they can just orchestrate these actions in any kind of way. That's what makes agents powerful. But now they can do it with more guardrails. But still, if agents can solve problems that you never anticipated, that introduces yet another challenge, and that is how do you build a UX for a task that you don't know at design time? Now think about it. You're building the application, there is a task you don't even know exists. How are you going to build the UX for that? And the answer is simply that you don't, because you can't. Right? And that is exactly why the future of UX is conversational. Because conversations are very flexible. You can throw anything at. Now let me tell you this here. That is not the future. That tiny chat box stuck in the corner of that screen, that is not the future in the future, which is happening now, we are flipping that model and we put the conversation front and center. This is your default ux. Now, does it mean that we lose all the richness that we got to love in the last few years and that we are stuck with just text? Well, of course not. But now these rich components are going to be generated on the fly in the flow of the conversation. And by the way, do you know how we are uniquely able to generate these rich components on the fly? It starts with an M, just giving you a hint. Metadata. Exactly. That is the secret sauce. Now, to deliver the future of ux, to deliver that conversational future, there is no better place than Slack, because Slack is already where these conversations happen. And now with agents, sorry, now with agentsforce, you can really add agents to any conversation, to any channel, which means that we can then collaborate with these agents and get Work done. This is the future of ux. And these are the three challenges that we are experiencing today. Agentic development, predictable agents, and conversational ux. Now, all of us in this room, we are lucky. And by the way, if you are listening online, you are lucky too. We are lucky because unlike other developers, we don't need to replatform because we work on a platform that's constantly evolving and that already supports these shifts. Now, this is exactly what we want to show you today. We want to show you how to embrace these shifts on the platform. Today we are going to talk about agentic development, including MCP servers, predictable agents and conversational ux. To get us started with agentic development, please welcome Stefan Chandler Garcia, Senior Director, Technical Content Engineering.

Stefan. Thank you very much, Christophe. It's so great to be here in the keynote room with our developers. Now, my favorite part of Dreamforce and any of these events is getting to talk to every single one of you over the last few days and really learning and understanding what you see as inspiration from our content, from what we're building, from our platform. And I feel like every three months or so I'm coming up on stage talking about how amazed I am at the shift in technology and the shift that AI has brought to us as developers. And I continue to see the same thing happening and so do a lot of you, Kate. Everything that we're building, everything that we're doing is changing thanks to these agentic platforms and agentic developer tools. We now have capabilities and tools that we can talk to, we can speak to to help us solve problems. They can even suggest changes for us. But we need to make sure that all of these tools have the power and support to maintain enterprise level security, design and development. Imagine if that assistant, that person you're talking to, maybe even a little rubber duck. Some of us knows everything about programming, knows everything about Salesforce because it's been ingested into its model. That is why we have introduced Agent Force Vibes. Agent Force Vibes is enterprise level vibe coding that you can use to bring into your business right into your salesforce. Org. It's a pair programmer that you have with you and not only does it help you develop these capabilities, are helping us as a company build features faster. Okay, you're going to see tons of stuff in this demo, whether it be agentforce, Vimes or features that we've been able to accelerate to market thanks to this amazing capability. So we're going to go ahead and check this out in a demo. We're going to start out by introducing you to Pronto. It's a brand new sample application that our team's been working on. I'm sure a lot of you are familiar with Dreamhouse and Coral Cloud Resorts. It's a tool that we're using to help developers get familiar with Theta 360 and AgentForce pronto is a food delivery company. They help deliver food for restaurants to customers. I'm sure many of you have done that this week. And inside of this app we have storefronts, we have restaurants, we have menus. I want to build a menu browser component for Pronto. Maybe something that we can distribute in the sample app. To do this, we're going to jump over to agentforce Fives. Now, I already know what I want to build because I've brainstormed, I've worked with an agent, I've worked with my team, and we've determined exactly what we want to show. We have a menu browser component designed here. We've already ran a plan here and designed that plan within agentforce Vibes, and we can go ahead and just run it. We're going to go ahead and ask agentforce Vibes to start building out this menu component because we've already specified what we want the component to look like, what features we want it to contain. It's going to go ahead and start running it now. One thing I want to point out though, is that I've already approved edit mode here, so that no matter what, it's going to go ahead and just start writing the files. If you want to have approval over every single change, you just have to untick that box and it's going to go through each one of those with you. But more importantly, here is I'm a developer and like most of you, I'm unbelievably opinionated. When I'm writing code, I know exactly what I want the code to look like. There are a lot of nuances in what I do and I can write those down into rules that I can store inside of my project. So I've wrote my own custom Agent Force five rules here around how I like my LWCS built. I use a lot of gridding components in them because I want them to be pixel perfect. And Agent Force 5 is able to ground on not only my style, but my team style guidelines. So if you're putting together these long, massive documents, which a lot of you probably have over the years, you're Agent Force five is context aware of all of those requirements that you have for your development. Now we Also have some that come out of the box. So if we check this out, we have standard Apex rules, app dev rules, tons of capabilities that we can use to bring into this. So our component spun up quite a bit. We can see it's still working along here. We've got just a little bit left to go. It's writing our Apex class for us. There's one feature that we've delivered in the last couple of months that I think has really blown everyone's minds more than anything this week, and that is our LWC local preview. It's been rebuilt inside of Agent Force Vibes so that as we're building these components, we can actually live preview them. So this is that component that was just built. We're going to go ahead and open up that menu browser right here in the Lightning Preview. There you go. That component we literally just Vibe coded with Agent Force 5 is available for us to use and view and iterate right here inside of our IP ide. However, I don't love the ui. It could use some changes. Thankfully, I have a great designer who's built a FIGMA file for me that we can use to actually ground on. I want to be able to take this design. I want to hand this over to Agent Force Vibes so that it can enhance my code, enhance my UI with the contents of this FIGMA file. So we're going to go back over to Agent Force Vibes and I'm going to show you. So let's go ahead and pass in a link to our FIGMA file and we're going to hit return there. Bear with me. Okay, and here we go. Let's go ahead and run our local preview here. As Agent Force 5 goes, I want to show you a couple of things here. What it's doing is it's taking our figma URL and it's going to download the layers from that document so that it can analyze them and provide analysis and feedback on the components that are part of that file. As this spins up, we're going to go ahead and we'll rerender the component in a second and we're going to actually deploy it into our org. You can see here, it's taking that JavaScript right here with us and it's rewriting the JavaScript based on that new design. In the same sense, it's also rewriting the HTML. Once that's finished, it's going to go ahead and show us the component. Let's go ahead and open that up here. In our local preview.

Perfect.

Instead we're going to go ahead and deploy it live. Let's see if we can deploy this right now into our Org. And then we're actually going to throw it right on the page ourselves. And there we go. It's identified the Org and the user that we want to deploy into. It's then going to actually attempt the deploy and it's going to push that right into our environment. There we go. Deploy succeeded. So now let's actually go back over to the Org real quick. I want to refresh my tab here in my Lightning app builder. Let's grab that fresh component. There we go. We have our menu browser ready to go. We just built it. It looks exactly like our Figma design. We're able to actually build that and Vibe code it in just a couple of steps. I know a lot of you are going to be iterating a little bit more when you're building your components and you can, you can continue to refine and build before you're actually deploying into your environment. So I actually want to take this time to announce one more thing. So what if you could use Agent Force Fives and all of this technology to build applications on any platform? What if you could build apps in your framework of choice, like React? We now can build and deploy applications in any framework on Salesforce. What you're seeing here inside of AgentForce 5 is a React app. This is a next JS app that we built for pronto inside of our SFTX project that's deployed onto our Salesforce infrastructure within the runtime with a session id. So in theory, you can build on just about any framework now inside of the platform. So what I want to show you is we will actually want to see If Agent Force 5 can make a change to this site. So again, inside of our live preview, so I'm going to go ahead and go to our restaurants page here. And what we want to do is we want to add a Star Rating component to this. So we're going to go ahead and open up Agent Force Vibes and we're going to go ahead and ask it to update the Star Rating component. And it's going to go ahead and it's going to make that change right inside of our component. Okay. And with that, we're able to actually make and modify our React apps that are deployed on Core right inside of our Salesforce projects. Now that's just a sneak peek at what's possible here inside of our applications on Salesforce. With agentforce Vibes, the possibilities are endless. And what you can build is just unbelievable, thanks to the capability that's now available. So I can't wait to see what you all do. Thank you very much.

Thank you so much. Stefan. Incredible demo. We speak about predictable in this keynote. I have to tell you, a month ago I didn't predict that we would vibe code, react in the developer keynote and then push that to core. But now let's speak about MCP servers because they are a big part of agentic development. To tell us more, please welcome Jay Hurst, SVP Product Management.

Thank you Christoph. Thank you developers for joining us today. Now, what we just saw from Stefan, I think is truly incredible. The fact that the teams have been able to pull this off in just the last few months in and of itself has shown how valuable AI can supercharge your engineers and their activities. So we've been working pretty heavily with agents and Agent Force and we found some of the same problems that you have. Specifically, we don't want just a smart LLM. We want agents that can be grounded deeply in our enterprise metadata, our schema, all of the customizations and capabilities that we've built over the last 20 years. This is where MCP comes in to take us through this and help us out. Because I'm not quite as brave as Stefan to drive my own demos, I'd like to welcome Aditya and Renee who will be helping us today. So the first thing we need to do, like we discussed, is bring all of that metadata together. Thankfully, we have a unified platform that will allow us to do this. And today I'm happy to announce that we're coming out with what we call the Unified Catalog. Now, the Unified Catalog allows us to take all of the customizations that you've done inside of your org, bring them into Data 360, and then build that unstructured rag and natural search capabilities against all of that metadata. So as we see, we can go ahead and search and look at any of the metadata that we have in our org. And we also have the ability to allow AI to generate these awesome descriptions that we may not have filled in ourselves. So we're using AI to help us build AI. Now, once we have all of this capability built up into the Unified Catalog, we want to get that into our Agent Force vibes to do the things that Stefan showed us. So we'll go next back into Agent Force 5. And the first thing we want to look at or talk about is mcp. Now, many of you have probably already heard about mcp. You've probably used it and, and for those of you that haven't, think of MCP as kind of a universal translator. It allows you to inform your agents and your AI where the tools that it needs to use live, what those tools can do, and most importantly, how to use all of those tools. Because that's the key. Just understanding there's a tool there doesn't help unless you can actually use it. So we're, we look up here and we see a number of MCP servers, the first being our Salesforce DX MCP server. Now this is all of the power of DX that the teams have spent the last decade building instantly exposed to your agents. Right? We can see all of the things that will be local to your development, like List all Orgs. List all Orgs is very useful when you're building against many, many scratch orgs. But that's not something that is hosted on the Salesforce platform. That's a local version and this is a local MCP server. But I'm also pleased to announce today that we are also releasing our beta of platform hosted MCP servers. So we can look up here and you'll see all of that rich API and service context deployed using mcp. So we can see things like our Describe globals query objects, create and modify your data, all hosted on the Salesforce platform. So you don't have to worry about building your own local MCP servers. We also will be adding the ability for you to create your own custom MCP servers and host them on our secure trusted platform with the rest of your CRM data. Okay, so we have our data localized into Unified Catalog. We've exposed it through mcp. Now let's take a quick look at how this works. First thing we're going to do is try and create a new tab for our rewards object. Now, tab creation is pretty straightforward. I'm sure all of you have done it at one time or another. But if you've ever used AI to try and do it before, it actually might take a turn or two to actually get that structure correct. That's because the AI only understands and has been trained on what's on the general Internet, not what's custom to your org. General Internet doesn't know what your rewards object is, but here we do because it's deeply embedded inside of that Unified Catalog. So we know what the shape of this should look like. Let's go ahead and deploy it again. Just by asking it through natural language, we can deploy. We see it's calling the MCP servers in the background. And just like that, we've been able to expose, create and and then deploy our components, all without actually having to write code, which I think is pretty fun. In fact, I would say that AgentForce Vibes has made coding that much more enjoyable. You can try things very fast, you.

Can.

Review, iterate and deploy instantly. Now we have one more little sneak peek. All of this power that we've shown is available inside of AgentForce vibes on your local machine, but we'll also be exposing it via the agentforce Vibe cli. So, quick demo. We can go ahead and look at all of our MCP servers, see a list of all of the ones that have been exposed, and you'll be able to use that same natural language processing directly from your cli. Now, MCP is great, we love it, we've been working very hard on it. But it's not just something we want to expose to you, we actually are deeply embedding it into our platform. Our internal development teams are using MCP to quickly bring the power of agents to their APIs. And to help demonstrate this, I'd like to bring up my colleague, Senior Director of Product Management, Karishma Lawani.

Karishma thank you J. It's truly fascinating to see how quickly we are able to create apps and deploy them using context aware intelligent agent tech technologies like McPeak and Salesforce wipes. But there is one more step you as a developer can take to make sure your application is ready for production and that is to make sure that you are scale testing it. Now up until now Scale center and Apex Guru were available to our unlimited edition customers only. But starting today I'm happy to announce that scale center and ApexGuru are available for all customer editions. And not just that, we heard your feedback about making Runtime Insights available early on in the build cycle. So I have one more announcement for you. ApexGuru Runtime Insights are now available in Code Analyzer in Agent for Swipes. As Jay mentioned, my team has been an early adopter in the MCP space. So today I'm going to show you a sneak peek into some of the pilot features that we are working on. For that we'll go back to a protorg. Now let's imagine Pronto has an upcoming championship weekend where they expect food delivery orders to surge. As a developer I want to make sure the application is ready to handle that load. So I run a scale test in my sandbox and here I am in Scale center looking at my performance posture of my org. But traditionally it is very Time consuming to correlate these complex metrics and charts. Not anymore. With Agent Force right in Scale center, this experience now becomes very simple and conversational. I can go ahead and ask my agent how's the health of my org? As you can see, the agent quickly summarizes which are the stable areas and where are the areas of concern. This is like finding needle in a haystack in a matter of sub seconds. Further along I can ask it to give me more details on where exactly are the hotspots. Again, the agent is able to pinpoint that there are some CPU timeout errors, roll lock errors. This is like having your own performance engineer right by your side helping you through the troubleshooting process. Now this is possible only because the agent is aware of all the orgs static metadata as well as runtime information that we are bringing into the Agent interface. Let's go ahead and ask how to fix these. Once again the agent quickly tells you what are the steps that you need to take to make these hotspots go away. This is how we are reducing time to detect triage as well as fix your code using Simple Agent in Scale Center. But that's not just it. I'll take the clap. We are talking about Multi UX conversational interface is the future. So for that I would like to go to my second demo. Now let's imagine pronto that championship weekend that we talked about is upon us. Here. Scale Agent is monitoring your production instance and as soon as it detects an anomaly it actually sends an alert to your team's Slack channel. Once again I want to pull in Scale Agent into the troubleshooting process to make this easy for me. So I go ahead and ask give me some insights on how do I fix these. Again in a matter of sub seconds, Agent is pulling all that information through our MCP server to identify what are the runtime insights here. If we expand the response we can see that it pinpoints to a specific class that is the problem which is triggering this alert. Now I will move on to agentforce Vibes to dive deeper into this class. If we look at agentforce Vibes, we've already configured our MCP server scale mcp so I will start chatting with the agent again. This time I ask it tell me where in the class is the problem Exactly? This is where the MCP server calls our Apex Guru tool in the backend to get information on the problem and it highlights this particular line which has an anti pattern called getglobaldescribe. Not only the anti pattern if you notice it's also giving you runtime analysis to say this is causing 35% of increased cpu. This is not just static anti patterns. We are actually getting production runtime insights early on. It also highlights another problem right in the class. This time it's 40% increase in database CPU. Let's go ahead and approve this. Once you hit approve the agent will now start showing you a code diff of what is the scalable code recommended. So as you can see it inserts a line and it says instead of how you've written it. If you write in a different way it expects the performance to be optimal. Let's go ahead and save it. What it does now it's actually making the changes of the recommended code that was suggested by Apex guru. Hit save one more time. Just like that the agent is done fixing code. Write an agent for swipes. In the end it will give you a summary of what was the problem, how did you fix it and recommended actions to avoid this in the future. Now this is how we are bringing everything from detection to causation to recommendation when all through multi UX interfaces. This is how we put enterprise grade scale into the hands of every developer. Back to you Christophe.

Thank you. Thank you so much. Karishma and Jay. Incredible demos. It's incredible to see how MCP servers can really extend our developer toolset. Now let's move to the next shift and this one was about how we make agents more predictable. And to tell us how, please welcome Avantika Ramesh, Senior Director, Product Management.

Amazing. Thank you Christophe. Hello developers. Good afternoon. Now ever since we launched AgentForce just over a year ago, my team and I have been on the ground with our customers like Equinox, Adeco and Pearson.

Hello.

Helping them build, test and deploy these agents. Now one of the common themes we noticed with our customers is that they wanted to prescribe very specific business processes within their agents. However, the responses from these agents were highly variable. Now of course LLMs and agents are inherently non deterministic, but when it comes to your enterprise rules and policies, even a 1% deviation from that is unacceptable. What you all were really looking for was the ability to inject specific moments of determinism within your agents. That's why my team and I are super excited to unveil the brand new reimagined and re architected agentforce builder. Now this agent force builder fuses both rules and reasoning to give you all a lot more deterministic control over how you build your agents. You can also leverage the help of AI assistance as you build to build faster. Now you have all the tools you need to and optimize these agents. Let's go ahead and take a look. So here we are in our Agent Force Builder. Now I want to quickly give you an overview of what you're looking at here. On the left we have our topics. This is the overall topology of our agents. And then on the right we have an AI assistant that's helping us as we build our agent. Now let's look at the center right here. Let's look at how Pronto uses this technology to build their order dispute agent. First thing they do, they have the agent name and description. Then we go into the topics, for example, order verification and escalation handling. And then finally they have their list of variables. Now let's dive into that first order verification topic right there. Now if we think about order verification, it's a very prescriptive business process. Essentially you need to make sure that when you're verifying the identity of a customer or or validity of an order, the AI actually follows that exact steps that you have specified. So here, as you can see, we're able to use natural language to set variables, define our conditional logic. And then finally we can also chain actions. So now you have a list of actions that can be chained deterministically versus relying on an agent to decide how to execute those actions. On the right here we have our AI assistant. So we're going to go ahead and see if our assistant can help us add an instruction. So let's say we have a high value order. We want to be able to verify the order through SMS through a one time passcode. There we go. Look at that. Our agent was able to add that instruction in line for us purely through this AI assistant. Now I want us to take a moment and compare this experience to what we have today in our Agent Force Builder. If you all recall today's experience, of course you can still build powerful, highly capable agents with what we have today. But we did hear you when you said that it became a little unwieldy to pack all this information into our instructions. So if you look at that, our instructions didn't have the ability to ground on your data, it didn't have the ability to reference variables. And when the agent executed these instructions it was based on a non deterministic protocol. It wasn't able to chain them in a deterministic fashion. So now we make it much easier to specify that level of determinism within your agents. So now that you see This I want to also show you another feature for developers specifically, which is everything here is natural language first, but as developers, you love to see the pro code view. And look at that. You can actually see the script that powers this agent. So we're actually going to pop over to agentforce Vibes. So here we are in agentforce Vibes, and what you're looking at here is Agent Script. Everything you build here is backed by Agent Script, which is Salesforce's proprietary language for agent development. From here for developers, you can see that it's extremely portable and customizable, and you all can do your magic as developers to make this your own. Now, what's the final part of building an agent? You build, you test. So we're going to go ahead and test this agent back in our Agent Builder. What you'll notice here is we actually have the ability to simulate interactions. Let's say a customer inquiry comes in saying that some, some of their items are missing from their delivery. Now you can actually see how the agent is going to execute all the different actions you specified within your topic. So let's take a look. So as this agent is reasoning, it's able to invoke that first order verification topic, looks up the order, and look at that, we have a response. It's able to summarize it, it's able to execute the exact sequence of actions we specified. And this is that hybrid reasoning engine in action. Yes, let's give that a round of applause. All right, so now that you saw all the innovations here in the AgentForce Builder, I want to take a moment to tell you all what my team on Prompt Builder has been up to over the last couple of months. So with Prompt Builder, we've been noticing that a lot of our customers, and by the way, we have close to 5,000 customers using PromptBuilder to power their AI apps and agents. A common pattern that we notice is that many of our customers were trying to write conditional logic within their prompts, if and else logic to evaluate specific conditions. But unfortunately, these models don't have a way to evaluate these expressions or apply logical precedents. So what we're doing now is we're actually bringing Agent Script closer to Prompt Builder for specific use cases like writing these conditional expressions. So as you can see, in this use case for pronto, they're actually able to write conditional expressions based on the delivery status and the payment method, and based on whichever clause evaluates to true, we're able to inject that specific instruction into the broader prompt in order to get our final response. So as you can see here, even Prompt Builder, you're able to fuse elements of expression based deterministic logic with the non determinism of an LLM. Overall, as you all saw, for both Prompt Builder and agentbuilder, we are combining both deterministic logic with predictable natural language logic. We're bringing deterministic models close to the probabilistic models and giving you all more control and consistency so you all can build those enterprise grade and most importantly predictable agents. So thank you all so much for following along. Ann, back to you. Christoph.

All right.

Thank you.

Thank you so much. Avantika. It's great to see a developer centric approach to agent development with with Agent Script. Now, the next shift was all about the future of UX and conversational ux. And to tell us more, please welcome Janani Narayanan, Senior Director, Product Management.

Thank you Christoph and hello developers. All right, okay, so so far we saw scaled agentic development, MCP power tools and predictable agents. Now how do you deliver all this goodness, all the intelligence and innovation you've built as developers into the hands of your users? That's what this chapter is all about. We're going to talk about conversational UIs now. With the help of Slack, which is the conversational UI of choice for us, you are able to build and stand up an end to end functioning agent in a matter of a few hours and we'll show you how. Now this is how you're able to take code to to collaboration and all that complex backend logic to an intuitive real time interface for your users. What is a highlight of this demo is that this is a real life Salesforce and Salesforce implementation that my team built. They're sitting somewhere here.

Woo hoo.

So it's not a concept, not a prototype, it's live in production and we're excited to show you the demo. But before I get to the demo, I want to give you a little bit of context on the use case that you're going to see. We're going to show you a demo of a paid search agent. Now what is a paid search agent? Now paid search marketing as a marketing discipline is one of the most mature digital ad markets out there. And I've been told that this is a $200 billion annual global spend market across enterprises. So it is massive, but also it is incredibly complex. Now our own marketing teams here at Salesforce, they have to manage thousands and thousands of keywords and campaigns and it's a constant game of iteration and optimization. Rinse and repeat. And at that scale, humans simply cannot keep up. And this is really where you have that big agentic unlock where you can automate repetitive tasks and constantly manage campaigns, as well as continuously bid on and optimize your bids so you can go after the $200 billion market faster than your competitor. Now that we have the context, let's jump to the demo. We already have the Slack ui. Now, it was obvious for us to choose Slack as a UI because that's where our teams are we're working on here day in and day out, collaborating. Now picture this. I am a paid search marketer at Salesforce and I have to figure out what are the top keywords to bid on to reach every potential Dreamforce attendee for my Dreamforce campaign. For instance, you sir, are probably googling what is the best AI conference to attend to, and this ma' am out here is probably looking for the best developer conference to go. She's not even looking at me. Now, the answers to all of this is pretty obvious, but we have to try, right? So as a regular search marketer, I am swivel sharing around five or six different dashboards. I have my Google Ads, LinkedIn Ads, Google Search Analytics. I also care about my SEO. So BrightEdge is a tool that gives me SEO analytics. So I am poring over all of these dashboards. It is an extremely manual and time consuming process for my team and I. But now I have this friendly little agent. So let me go ahead and ask the agent what keywords should I bid on for the upcoming Dreamforce campaign. Now, as the agent is doing its thinking, let me unpack what's going on behind the scenes. The agent simply doesn't see those words. It understands context. It sees me as a Salesforce user, understands what permissions I have. I have all of those campaign details handy as well. Those are tabbed away in the Slack UI. In addition, all those aforementioned data sources, Google Ads, LinkedIn Ads, Search Console, Google Analytics data are all available in Data360 via Marketing Intelligence Connectors. So the agent can plug into the data, compile a list of keywords and recommendations, suggest ad groups, and publish the results right here on Slack UI very neatly in a Slack Canvas as. So there you go. So what took me hours and hours, even weeks in a matter of a few seconds. You see the list of keywords, impressions, clicks, all the insights that my team and I can use. Great. Now that we've seen this through the lens of a paid marketer, let's take a look at what's happening under the hood. Let's stop over at agentforce Builder. All right, now the topic that gets picked up for the response is keyword research that then executes a sequence of actions such as Create Slack Canvas, Update Slack Canvas search slack canvas, etc. The point to note here is that all of these Slack actions are out of the box available right here in the builder. Slack connection is also native to the builder. So as a developer you can simply chain them together in your logic. And you're already operating on the Salesforce Trust layer. So things such as permissioning, auditability, real time updates are all available to you right here. So extremely powerful. All right, so let's switch gears a little bit and now I'm going to show you a Tableau Nets dashboard. Now if you recollect I said this is a real life implementation at Salesforce, our marketers use the agent and they also have access to this specific dashboard so they can get additional insights about performance per channel and other top line ad metrics. Now the point to note here is that the data that powers our agent is also available to the Tableau Nets Dashboard. And what makes it possible is that shared semantic layer. What you're looking at at the screen is the custom data model object, semantic data model object that my team built to power the agents as well as the dashboard. So what does the agent do with it? When the agent encounters a phrase such as campaign performance or any ad metric, it is not doing any guesswork, it is putting pulling in from the definitions that are codified inside the semantic data model object. So as a developer, you can query it, extend it, embed it, and you're not starting off from raw SQL or field mappings. You are pulling in from the governed data ontology layer. Very powerful. All right, so one last stop, we're going over to Prompt Builder. This is a brand new prompt builder that Avantika's team built. Now this prompt template that you're looking at, if you remember, my search results were on a Slack Canvas. It was tabular, it was neatly structured. So this prompt template allows me to take the semantic query results, which is in a JSON format, and then convert it into a Slack Canvas compatible markdown JSON format. So it acts as the translation layer between my semantic query results and my Slack experience. So as a developer, you simply instruct the model once on the structure, you want to see it rendered and then it is reusable, composable, and no additional custom code is needed to format your insights.

Great.

So we saw a lot here. We're not quite done with the demo yet, but I'm going to give you a quick recap. We saw the agent in action. We saw Slack as the ui. We also saw Agent Force Builder that came and shipped with out of the box Slack action. Slack was native to the Agent Force Builder and we showed you a tableau Nets dashboards that our marketers use and the semantic data object model that powers both the dashboard as well as the agent. Who's ready to see some code? Come on you guys. Developers. We are on Agent Force Vibes favorite tool of the week. Now as I was conversing with my agent, it was giving me text messages. But every single response is backed up by a semantic query. What's cool about it is that this semantic query is executed via the Semantic Gateway API. So I can simply go grab that query and to demonstrate that I can simply run it out of Apex here. So as Aditya is hitting run now, the response here, the exact same response that you would see on the Slack canvas. Identical content in the same structure. So the big takeaway is that if you can run it out of Apex, you can run it out of anywhere. And to bring home that message, let's go back to our food delivery app that's made me hungry throughout the keynote. All right, so the charts and the tables that we see here is also backed up by the same semantic query. So the big takeaway, Semantic queries are portable. They are API first objects and you can execute it out of multiple surfaces. Apex node, Python, your favorite custom app and it always resolves against the same governed data model object. Very, very powerful. Now that brings us to the end of this demo. As you saw, even in conversational UI layer you see that inflection point where the future is not quite about just building apps. It's probably about building intelligence for your users to interact with, reason with and perhaps even talk to. And with that little nugget. Back to you Christophe.

Thank you. Thank you so much Janani. Incredible demo. These are the shifts that we are experiencing today. And now I would like to invite someone who has been on that journey for some time. Please welcome Maham Asan. Wow Maham, you have a fan club. Welcome. Welcome to Dreamforce.

Thank you Christoph. So hi all. Maham Hassan. I am senior solutions architect at Conix Digital. We are the leading partner in the Middle East. I'm a Salesforce mvp. I'm a proud community advisory board member running a developer group in Dubai and just love everything which is community.

That is totally awesome. And we've been talking about agentic AI and agent 4 scale. Can you tell us a little bit about your journey and any advice that you have for people who are just getting started?

100%. My AgentForce journey is a reflection of how basically you turn your curiosities into your milestones because it's been a year since I've seen a huge amount of this product. So we at Connex have built multiple agent first POCs and we are moving them into production. So that's a very exciting time for us. And based off of my experience, everyone in the room who are going for their first agent force implementation, I'd like to recommend the same structure that I use for my agent force implementation which starts from conversation, planning and execution. So I usually start from deeper discovery getting involved with my customers. In my case, Christoph, I have like sat with my stakeholders customer service team for hours and hours putting my headphones on and sort of understanding that how they are resolving their customer issues and how they're facilitating their customers and so that I can sort of create the same blueprint onto a digital laborer to help them do their jobs better and empower them with data and AI. The second step is definitely planning your agent. So there is just like in these two days we learn just so much about how you can build your agent and how many ways are there out. So planning comes in place. So either you are coming up from no code to pro code agent or let's say deterministic to probabilistic. You need to ensure that all of that solutions are aligning with and addressing the use case of your customers and how basically you're going to solve them with the digital labor. And the third thing is whatever you have planned, stick to it, deploy it, no, implement it, test it, deploy it and then make sure that the KPIs are fulfilled.

That sounds like a great plan. Now with the community, you're never alone. So can you tell us a little bit about your journey with the community?

Sure Christoph. I started my journey, my developer journey 30 years back. All these cool tools were not there. So you can definitely call me a byproduct of stack overflow maybe right? But yes, today I'm here because I've seen the power of this product. It has made me an advocate and what my mission is to basically empower my community and to bridge the gap between the potential of this product and the power of my community. So technically through the developer group that I'm leading and the by training, we are just mentoring people, we are running workshop, we are in the next month, like at the 15th of November, 2025, we are doing our second biggest conference of the Middle east. And we are empowering and mentoring 50 trailblazers and agent Blazer to become Agent Blazer Legends.

Wow. That is incredible. Really inspiring. Maham, we appreciate everything that you do with the technology and the community. And we have a little token of appreciation for you. Please join me in congratulating Maham on her golden hoodie.

Thank you very much.

Congratulations.

Thank you so much.

Congratulations, Maham. Thank you so much.

Thank you very much.

All right. Thank you.

All right.

All right. And now it's time for all of us to embark on that journey. We have plenty of activations downstairs to learn. And after Dreamforce, join the agentforce now workshops to learn more. Thank you so much. And have a great rest of Dreamforce.

Thank you.