The Enterprise AI Harness for the Agentic Enterprise | Dreamforce 2026

Please welcome President, Chief Platform and Engineering Officer, Salesforce, Rohan Kumar. Hello, everyone, and welcome to Dreamforce. You know, I've been here about 3 months, and I've spent the last 3 months actually talking to all our customers, many of them across the world. I've just been traveling quite a bit, you know, so I've met CIOs, CTOs, CDOs, security leaders, and business leaders and practitioners as well, you know. And the one thing that they all sort of say is, hey, this is not just— this whole AI transition is not just a change in the AI stack, it's actually changing every role in the enterprise. And that's pretty amazing. So we basically want to look in in this session hear what's the thing that we are building to enable these transitions to happen successfully. All right, but before we do that, I want to take a moment to thank all of you. You know, thank you for trusting us with your data. Thank you for trusting us with your applications. And thank you for trusting us with your AI. It means a lot, and we take that trust very, very seriously. You know, you saw Mark in a keynote this morning speak about the new stack for the age of AI, starting with AI Force, which is the experience layer. That's how you're going to experience AI across various functions. Agent Force essentially is where all the agents and the runtime— agents get built and run. The Customer 360 is essentially all the applications which are the backbone of collecting all the information and getting the business workflows executed. And then Data 360, which is the foundation of how all of these things come together and just to get your data right. And it's interesting, like, if you take a look at the transition we are going through, every enterprise, and this is if you just take a look at the last couple of decades, was built around people, which makes a lot of sense. So you had people opening up screens, looking at permissions, workflows, processes. They clicked on applications. Right? They essentially opened up CRM if they wanted to deal with anything that was front office facing. You know, they opened up analytics applications if they wanted to do BI, etc. And the fundamental part of this change is we are moving from an application-centric, you know, era where, you know, humans basically opened these apps, clicks, you know, clicked in a bunch of places and got things done to an intelligence-centric era. And what do I mean by that? Instead of doing all the work yourselves, like the practitioners are essentially going to ask the agents to get work done by them, right? That essentially is a fundamental change. So clicking of the screens essentially is going to transform with creating these prompts, creating these loops to actually get the work done. That's actually the fundamental change. Now, in the conversations that we've been having, There's a few key themes that have come— that are coming up. The first essentially is customers are seeking outcomes, not just AI for the sake of AI. You know, everyone's basically tired of token maxing. They've just done a lot of work and there's not a lot of value to sort of show for that, right? So they are seeking customer outcomes and not just, you know, AI projects. I mean, this is a great example, right? So if you have an agent that essentially takes some input and then keeps asking questions or escalates it to human, That's not very valuable. But if you have an agent that deeply understands you as a customer, right, has the entire context, like what's your business, what is it that you're trying to accomplish, and then when you call for a support scenario, it's able to actually get all that context in and solve that without escalating to a human. Now that's real value. That's a very specific customer outcome. And that's the key thing that we are focused on. Now, the other big learning is the agentic enterprise changes every role. So it's not just the business user. I mean, you see a lot of focus on agents from a business user standpoint. But the reality is one of the things we have learned is every role in the enterprise. And let's take a couple of examples. If you take a look at the role of the IT leader, the CIO, right? They've managed applications, machines, users. And now with this digital workforce where you have all these agents that are running around in your teams, inside your platforms, inside your applications. The big thing is how do you sort of manage, govern these agents? How do you look at things like cost control? This thing is becoming extremely important. Like, what value am I getting out of these agents? That's going to become the job of the IT leader. You think about the security leader in the company. They've managed, of course, same thing, user identities, devices, to a certain extent, data. But with the agents, the whole game changes. Most of the agents today are run— running on behalf of a user identity, and that is very risky because the users have a lot of privileges, right? And your data is not very organized. So the IT leaders are significantly looking at, hey, how do we transition to this agent identity thing where the agents have their own embodiment and can be controlled and governed like that? If my data is not classified properly, if my data is not governed properly, That's a huge risk because if these agents go rogue, they can leak your confidential data. So this is how all these roles essentially are changing. It's just a couple of examples. The other thing we've learned is enterprises have always had fragmented stacks. This has always been a challenge, and it just gets exacerbated with AI because you have to get your data right. And there's a lot of focus that's been there in enterprises to connect all these various systems to normalize the data and get it right. But just connecting these systems is not enough. You need to pick something which is a center of gravity, like what's sort of the foundation around which you're going to change your enterprise. And that has to be the customer. This is the other big piece we've learned where when you're evolving, when you're evolving to become an agentic enterprise, you have to think about the customer. Make that the centerpiece of your strategy as you're thinking about this transition. And so that's the other big learning that we've had. And to take a look at, okay, how do we solve all these challenges? Fragmented stacks, focusing on the customer, really thinking about customer outcomes. How do we look at this? For that, let me introduce Des to come and give us the idea.

Thanks, Rohan. So if the agentic enterprise is organized around customers and the business outcomes they want to deliver on, The question we need to ask, how does AI make that possible? And the reason behind that is that models are incredibly intelligent. They can reason, they can plan, they can act. But models are also probabilistic, which means they give you answers that are likely to be the answer. And in the enterprise, you need to have answers that are grounded in truth. They need to have— make sure that it's the right answer every single time. Because models don't inherently know your, your business. They don't know your customer data. They don't know your workflows, your policies. They don't know your governance. And even though the model is gonna get more and more intelligent over time, doesn't mean that intelligence can run your enterprise. They need something around it. And that something is what we call an AI harness. Now, the AI harness is everything that surrounds the model that makes your enterprise work. It gets all the context from your organization. It gets all your workflows, your policies, your permissions, so you can act with trust. And the best way to think about this is that the model brings the intelligence and the harness brings your business. But let's actually talk about what's driving this transformation. And it all starts with the model. Now, we have lots of different model choices, as you can see. My model of choice is Claude Code. I use it every single day. And as I mentioned, models can actually analyze, they can plan, they can create. You should see what my kids are creating these days. It's insane. But models inherently don't know your business. And your business are all the things that you see here on the next slide here, which is everything from your customers, your data, your workflows, your applications, your governance, your institutional knowledge, your history. Now, a model will know what an order is, but a model won't know the customer that's associated with that order. And a model won't know that that customer has actually called in 3 times in the last week saying, I haven't got that order. And the model won't know that that inventory for that order is actually stuck in Italy right now. Because at the end of the day, the model needs to be connected to your enterprise. And the way that we do that is where the AR harness comes in. It now is able to get all that business understanding, your data, your context, the customers, the history. It can also now connect to your agents, your APIs, all of your different systems so it can act on your behalf. It also has all the control with governance and security and trust and cost controls that you can work within the enterprise. And once you have these things connected, that's when you start to see real business outcomes. You're able to grow revenue. You're able to delight customers. You're able to reduce costs. You start to see that increased productivity. But more importantly, you can move even faster now with these business outcomes at your fingertips. And I want to actually bring someone on stage who actually is seeing a lot of this growth right now in the enterprise. I want to introduce Scott White from Anthropic to talk a little bit about how you're seeing this today. Welcome, Scott.

Thank you. Thank you for having me.

Great. Is this your first re:Inforce? Have you been before?

I've been before.

Okay, wonderful. Well, we know that you've been talking to a lot of enterprises, and I'd love to hear what are you seeing the ones that are really successful and seeing the full value of this, and if there's any advice that you can give to the audience today.

Yeah, sure. So I think the first thing is probably the foundation. You've talked about this. It is, you know, an agent can't actually do anything for you if it's not connected to your enterprise data and doesn't have the ability to take action in systems. So the first thing is connecting into the right data sources to be able to take those actions and creating the right tools so that the model and the agent harness can actually take very significant actions to get to business outcomes. But to be comfortable deploying agents that can take action and use context, you do need to build the right guardrails. What are those tools? How do they map to permissions and your role-based access control within your organization? which actions can be taken without a human approving it, which have to have human approval. And then after all of that happens, you know, what are the actual outcomes that you're driving when you're connecting multiple workflows that even might be disjointed in nature? What are the business outcomes that are happening? What are the areas for optimization? And it's very exciting because when you do have that foundation, you can really unconstrain what I think of as sort of frontier trailblazers. There are people— we've seen this— there are people who really a small group of people in an organization can do really transformative things and sort of be the model and the sort of paradigm by which other people look at and say, okay, we can do really ambitious things and we can do more. So it's about finding and unconstraining those people with the right foundation in place that can then scale the best practices to your organization. Because with the right observability, you can find out what makes them different and then package that into behaviors to through agent skills and connectors to really raise the floor for the rest of the organization as well.

That's awesome. And now that we've been talking about this harness, imagine now that we can connect the model intelligence with this harness. Tell us what you see as these coming together.

Yeah, absolutely. Well, I mean, first I think of Claude as a sort of reasoning layer that can orchestrate across multiple different systems. And then Salesforce brings this sort of enterprise harness in terms of the connectivity, the data access, the governance, and the tools to allow agents to actually take meaningful outcomes or actions that lead to outcomes. And, you know, when you combine these 2 things, I think I've seen 2 things happen. The first is that you give people what I think of as sort of professional superpowers, right? Like doing something that is in their role, but doing it, you know, 10 or 100 times more effectively. And the second, which is I think really exciting, is you start to see people expand outside of their traditional purview. And if you think of maybe just an example like, getting a renewal done. Imagine you sit down and you talk to Claude about it and you say, I want to get this renewal done. Well, Claude can connect into your data warehouse to actually look at and understand what are the unique customer behaviors, what does the actual data say. It can look at your Salesforce instance where there's a rich amount of information about that customer, your Slack conversations where people are talking about the use cases and the opportunities to grow the account. And then it can say, okay, I'm coming up with a hypothesis about how to structure this. and then start to do work for you that you might not have done before. Create the deck for the proposal grounded in that customer data, you know, actually maybe kick off a redline proposal that you can give to your legal team and say, this is my opinion about what they want to do and where, you know, the ambitious thing that I'm trying to do with them. These are things that maybe would've been traditionally outside of the purview of a seller before and require a lot of handoffs for the data team to figure something out, the ops team, the legal team. And now they're really doing it themselves by orchestrating sort of enterprise AI harness. And I'm very excited about what that means in terms of people expanding outside of their traditional purview.

Awesome. I think you really landed why the AI harness matters. So thank you so much, Scott. I really appreciate it. Welcome.

Thank you.

Thank you. All right. Well, now that we kind of gave you a little sneak peek into what the harness works like, I'm going to bring Rohan back on stage to talk a little bit about how it works within Salesforce. Back to you.

Thank you, Des, and thank you, Scott. That was wonderful. All right. This is probably the most exciting part of Dreamforce for me to actually walk you through Salesforce's Enterprise AI Harness. It's, again, sort of coming back to the stack that Mark showed and the evolution. Where this harness is starting off is what we have with Data 360. We are essentially taking Data 360 and evolving that into what we call the Salesforce's Enterprise AI Harness. Now, just at the top level, if you sort of step back, there's 6 capabilities of the harness, and that's wrapped together by what we call the AI control plane. And I'll quickly walk through them. Trusted models is all about choice. You know, one of the most consistent pieces of feedback that we heard from our customers is, look, we want the choice when it comes to our AI, right? We wanna pick the right model as measured by price performance for the specific outcome that we are seeking. It's extremely important because that's the way in which you can sort of control your costs and get the work done. And today, we're very excited to announce COA, which Mark did in his keynote. It's essentially the model, the first CRM reasoning model that Salesforce has released. And we've built this on top of Nimetron. And it's pretty exciting because it's 27 years of product research from Salesforce, which has sort of gone into, from a value standpoint, to making this happen. So pretty excited about it. But again, it's not just about our model. It's really about the amazing amount of frontier models. Of course, Claude is huge, and we want to make sure that the customers have a choice over there. The second thing is around trusted context. Think about like, this is really where all the business understanding of your company comes together. Like very, very critical part of the harness where we reach into these multiple data systems that Des and Scott were talking about. synthesize the enterprise knowledge. I mean, this is something that Salesforce is really great at. We've done this with CDP, identity resolution, and so bringing all those capabilities together to create the context. The third piece essentially is agency. This is really where when a customer expresses an outcome of this is what they're desiring, what are the steps that the agents, the plan that it creates, the runtime, the evals, all that's captured as a part of trusted agency, right? So this is Think about all the decisions that the agent makes to get a specific outcome done. The 4th thing essentially is what we call trusted actions. Now, once you've decided, once the agent has come up with a plan, a set of decisions that needs to take, you have to execute those decisions in a secure and governed manner. Does the agent have the permissions to go update the CRM or not?

Right.

So that's the reason why we are carving out actions as a capability by itself, outside of the trusted agency piece that I spoke about. And finally, trusted governance. This is again really about getting your data right before you sort of get to your trusted context. Think about all the data right policies, classification, etc. All that is a part of trusted governance. This is foundational. Having the right trusted governance is foundation to creating the right trusted context. And finally, the last capability is around trusted security. So think about things like agent identity, right? Based on the actions, the observability that you have around agents, You may determine that a specific agent is more risky than the others that you see. There's a certain threshold, and getting all that observability data, flagging that for the agent, all that's a part of trusted security. The other part essentially here is your data security. Data which sort of backs your trusted context is the most important asset that you have in your organization. So you have to protect it. You have to classify it. You have to have policies around protecting it. What can agents do with certain things that are marked as highly confidential? So things like that are all captured inside Trusted Security. The one important thing essentially is while all these are distinct capabilities, they're all brought together by what we call the AI control plane. This is in some sense the control tower that the IT leaders are going to use to manage all the agents that run in their enterprise. Starting from discovering all the agents, a single registry with— and it's not just agents that are built inside Salesforce, the Agentforce harness. If you have built agents in Microsoft and Google and Amazon, in any of the frontier model providers, all those will be discovered in real time and cataloged inside our agent registry, right? And then there's the governance piece of it. How do you sort of do cost control? How do you essentially understand the performance of an agent? What's the value that you're getting from the agent? All those details will be captured as a part of the control plane. Now, stepping back, while from Salesforce, we are envisioning this Enterprise AI Harness to be a complete solution, right? When we think about— we've spoken about the headless APIs, right? All this is accessible via headless. And it gets the enterprise to where your AI is. That's the vision if you imagine. If you can wrap around your hands around the harness, you should be able to bring it wherever you— for the work that you're trying to do. But we also know that most of our enterprises just don't have Salesforce. It's very complex systems. You have multiple vendors. So being open and composable is a very, very critical strategy of the harness, where if you have a different vendor for security, a different vendor for data, a different vendor for your governance of your control plane, as an example, that's totally fine. You basically just use the capabilities that you need from the harness. And through our headless APIs, we will integrate with all these. And we're going to build a very rich partner ecosystem around this. So open ecosystem composability is a very, very key design principle that we are approaching this with. And finally, you're not starting from scratch. In case you're wondering, all this stuff is great, but I've invested in so much of Salesforce. So if you're using any of our cloud sales, service, marketing, et cetera, using any of our platform products, be it MuleSoft, Informatica, Tableau, Data360, our core platform, any of these things, you're already on the journey to be using our enterprise AI harness when we have this. Because we basically, we are decomposing the capabilities of these products and bringing them together as a part of these 6 capabilities in the control plane that I spoke about. So just as an example, I'll tell you trusted context is capabilities from Informatica, from Tableau, and from Data360 and the core platform. We are breaking them down, and we are assembling them into this amazing enterprise AI HANA. So enough of me talking about it. I want to invite Jeremy. Jeremy, please. I mean, CTO of AT&T. First of all, Jeremy, please, come on.

Come in.

Great to see you.

Thank you.

I'm feeling very energetic today.

Thank you. There's one thing about me talking about this stuff, but you've been a Salesforce customer. Thank you so much for that. And very recently, we've actually had a project with Agentforce that's gone live. Can you share a little bit about it?

Sure. Imagine— how many of you have done a phone upgrade? How many of you enjoyed the experience? Fewer hands. So what if in 8 weeks, you could take 2/3 of the time out in doing a phone upgrade, go from 30 minutes to 10 minutes inside of a retail store, and actually create an app on top of a headless Salesforce and deliver that live in that amount of time. That's what we just did.

Wow. That's wonderful.

And so a great partnership with Salesforce on this. We literally did this in that amount of time. It is now actually already live in our retail stores. We'll be scaling it across all 5,000 retail stores here over the course of the next number of weeks. And it really fit in nicely with the themes here. It's a React-based app. We can— it's in a headless environment. We can upgrade and update that app as we need to, and it plugs straight into the Salesforce implementation that we already have and a lot of the other complex systems that we have, and we're leveraging MuleSoft for that. So it's been a great success.

Wonderful. Thank you, Jeremy. And I think it's amazing because partnerships like these where we work closely with you and sort of land this Really helps us improve our product. So thank you so much. It's wonderful to see. And yeah, that was record time for sure.

Yeah, that doesn't happen very often. You know, we at AT&T, we have been taking customer orders since the 1800s, just to give you some context. So, you know, moving at that kind of speed is pretty remarkable for a company like ourselves.

Well, thank you. So, you know, you've heard us talk about the Enterprise AI Harness, and I know we've had a couple of meetings on that. So when you just think about the composability piece, the data and AI, Like, what are your thoughts? Are we on the right track? Any guidance as we are working on this?

I think we're very much on the right track. This space is evolving obviously incredibly quickly. But if you take an enterprise environment like AT&T, you're talking about 6,000 or 7,000 different systems that have— some of which have existed for long periods of time. Some of them are extremely complicated systems given all the different lines of business that we're in. So the ability to leverage not only just the headless architectures, but also from an AI perspective, being able to use any model we want, Being able to use any database that we want, it allows us to actually insert this stuff into our technology stacks much more quickly than traditionally we were able to. So as you think about, you know, whether it's SaaS providers or other providers, things were more prescriptive historically. You had a UI/UX you had to have. There was a database you had to have. Now we're actually able to insert that into a company like ourselves in record time. So we're not only doing what we did with the application, we're doing this now leveraging the AI components from AgentForce. all across our technology stack.

That's wonderful to hear. I think it's very, very aligned with the vision that we just laid out. And ladies and gentlemen, Jeremy Legg. Thank you so much.

Thank you.

Wonderful. Thank you so much. All right.

That is—

it's always so fantastic to hear from a customer like AT&T. I mean, just the scale is— it sort of always blows my mind. Now, with that, let me invite MK. Where's MK?

Oh, there you are.

All right.

Hey, Rohan.

So actually, so we have a visionary demo planned for you. Obviously, this is sausage making in progress, but we do want to show you what our vision is over here so you see where we are going with this harness. I'm really excited about it, but we'd love to hear from you guys.

Thank you, Rohan. Are you guys all ready to create the agentic enterprise? All excited?

Yeah.

But you know what? Beneath that excitement is a hard reality, which is you can't just give the model— however powerful the model is, just throw some MCP and some data on it. And expect it to get your business right. You really need a harness that is pulling all of your enterprise knowledge, your content, your workflows, your actions, and security— pretty important— together to give that power to that model so it can give you the right answers and the best answers. Now, enough of me talking. We want to actually show you a demo of it. It's going to be in 2 parts. Gabe's gonna do the first part. Gabe's somewhere there, and he's gonna do the first part where we are gonna have Maya. Oh, there you go. Maya, the person who is a sales representative, she's trying to get her Q3 numbers right. We are gonna show before, after, without the harness and with the harness. And then I'm gonna be doing part 2, showing you behind the scenes of how this is all gonna come together. You ready, Gabe?

I'm ready. Thanks so much, MK. So we're looking at Claude here, and we have equipped Claude with a bunch of our enterprise data so that it can help our employees work. And as MK just said, Maya is a sales rep, and she has a hard question for Claude. She is a bit behind on her Q3 number, and she wants Claude's help to figure out what customers she should focus on to actually close that gap. And while Claude is thinking, Let's talk a little bit about why that is a hard question. To answer this question, Claude needs to understand what Maya's Q3 goals are. It needs to understand what customers might actually qualify to help her close that gap, and it needs to understand what Q3 actually means inside of our enterprise. And if I scroll below, 43,000 tokens later, Maya has an answer. That is wrong. And it's wrong because it's showing a frozen customer. It's flagging a renewal that's already in the pipeline, and Claude doesn't understand what Q3 actually means inside of our enterprise. And that fundamental lack of understanding about our business, it pollutes the entire thought chain. And none of this is Claude's fault. It's not Claude's fault because I've equipped Claude with data, but not actually an understanding about my business. So what we're gonna do here is we're gonna try this again, but this time what we're going to do is utilize a skill that we've installed here, and we're just gonna use that skill. This teaches Claude how to talk to the enterprise, um, AI harness, and we're just gonna re-ask Maya's question. And now we're gonna see something very different because Claude doesn't just have raw data about the enterprise. Instead, we've equipped it with trusted context. And what does that mean? Well, instead of the raw data, what Claude can do here is actually tap into the actual sales pipeline from Sales Cloud. And if I scroll below, you'll see Claude is able to use the semantic layer to get an understanding about what all that data actually means. That includes what Q3 means inside of our enterprise. If I scroll a little lower, we'll see Claude is actually able to tap into memories. These are observations that have been made about the customer from actual customer interactions. That we want to preserve across all interactions to give deeper context about the customers. And then it's able to take all of this rich data, compare it to all the rich data that we have across all of Salesforce about the customer, as well as external data from other systems to produce a much better answer, an answer that is now grounded in all this context. And if we look, and, and it's actually doing this while spending less tokens. And that's because Claude is no longer struggling to kind of interpret raw data. It can just tap directly into that customer context. And it's telling Maya that she should focus on Northwind and why. And if we look a bit closer, it's also flagged a condition here because this plan only works if this delivery can be made in Q3.

Mm-hmm.

But because of the enterprise AI harness, Claude is also aware that it can help Maya actually validate that this is possible. Now, while Claude is working here, this is where we get to see something else Rohan was talking about, which is trusted agency. We want AI to be capable, free to help us pursue business outcomes, but do it safely. And what Claude has been able to do here is tap into the AI control plane and actually discover specialist agents that it can work with to actually validate that this commitment is possible. And then it's been able to take all the security and all that context we saw earlier and orchestrate that work across all these specialist agents to ultimately produce another great answer that, yes, indeed, we can honor this commitment. But if we look at kind of the Tableau dashboard that has been embedded here, we're getting a hint that there is some maybe delivery and sustainability risk. But here again, Because of the enterprise AI harness, Claude knows that it can also help Maya actually execute all this. Now, this is where we get to see trust in action. Because instead of just giving Claude raw access to the database to just update, what we've actually done here is equipped Claude with actually approved business processes here across the organization that Claude is going to use to actually execute all of these actions. And if we look close, we'll see that it was not able to actually execute one of these business processes. It ran into a security rule, and it— the agent that it's using behind the scenes doesn't actually have access to grant this, to actually complete this action. And Maya, in this case, does. So she can take her authority and transfer it over to that agent. to get an outcome that every enterprise is ultimately looking for, which is AI that is trusted. But the big question behind all of this is, how are we going to help you create an enterprise AI harness that enables you to create this reality inside your organization? And to answer that question, I'm going to do what MK said, which is I'm going to turn this demo over to him to show us what the team has been building. MK.

All right.

Thank you, Gabe. And did you also notice it was far less token burned there? It was 43,000 to 7,000. Yes, that's what context can— the right context can make that happen. All right. So quickly, you're going to see how behind the scenes all this came together. Remember, with and without the harness. And this is the harness that you're going to be— that you're seeing here. So you're welcomed by the screen, which is a coworker screen. I'm going to first see what my connected systems are. With the Enterprise Harness, we can scan all of your systems, all of your enterprise systems, bring together all of your data systems that are out there, quickly know what they are, all the models that you're using, the gateways, of course APIs, and you can convert any of the APIs to MCPs with a single click. And finally, all the agents that are running in your enterprise. These could be Agentforce agents, Anthropic agents, things on Google, doesn't matter where they are. So that is the first thing. We have discovery right out of the box where we can understand everything that you have in your enterprise. Now, remember what we talked about, what Gabe talked about and Rohan did, is about the context we want to give to the agent. That starts with data. Given we have access to all the data, we know everything that is there in the enterprise across all of your data assets, whether it's from Salesforce, Databricks, Snowflake, doesn't matter where it is. And we can quickly look at the data. If you want to clean the data, transform it, you can do it all in place. We can also attach the semantic meaning to the data. What does sales order mean in your thing? Or in that, in the example he talked about, what does Q3 mean? Like the fiscal Q3, it ends basically October, not September. So those kind of glossaries, those kind of semantic meaning of that data, you can also create. Now, this is all structured data. Not just that, you can now index any of your unstructured data out of the box. All you have to do is just give a link to where your unstructured corpus is, whether it's in SharePoint, whether it's in some files, Google Docs, doesn't matter. We can now index, semantically index it, but Beyond semantic index, we also can do knowledge mining out of that automatically. You don't have to say anything, and we can automatically infer the graphs from all those unstructured documents. That means even if somebody in a call transcript had referenced some company name or some deal name, we can now link all that in, in a graph. So I can quickly zoom in on the graph, look at the schema, and see how these documents are all coming together. That is creating you that data context of your enterprise. All your data, all the other metadata, structured, unstructured, comes together. Now you got to make it easy for all these agents to access it. So what we have built is this new thing called a content context pack where you can simply say, you know what, for all my sales agents in English, plain English, you can just describe this is what they probably need to get access to. Because you don't wanna just give— overwhelm an agent with like 50,000 different documents from like 2005. You wanna give it, you know what, go use this particular subset of it that makes sense for you. All you have to do is just describe and the system will automatically figure out what tables to pull or what things to bring in. And it even compiles it into the right SQL, SOQL, GraphQL, whatever it is, or APIs that you're going to need. That's all you have to do. Get your data ready, get your context ready, and now you can go create your agents. Now we have out-of-the-box agent builder experiences with our powerful Agentforce platform where you can go build sophisticated both probabilistic and deterministic agents. Determinism plus probabilism is a very critical thing. You, you have some business processes You can now control exactly what happens before the reasoner step or after the reasoner step. You have full control of that. And of course, you can now drop in all those context packs right there for the agent to be able to use it. Now, what if you have agents not on the Salesforce platform? What if you want to go build it on other platforms? Not a problem. With just 3 lines of code— or rather, 10 lines of code here— you just have to drop into your Pro Code agents. you can use the same context you created, whether it's a Google agent, a LangGraph agent, a Microsoft agent, doesn't matter. So you can use this context with our agents or your agents, and the system will support both. All right, now let's go back to that unified view where we saw all those agents. Remember this customer commitment agent that Gabe showed you? This was the agent that was basically going around and like going into all the other agents and getting the commitments done from SAP and other agents. So how did it actually get all this live? So we can go ask Coworker that question. How did it get all that context? Well, guess what? First, we had these context packs that we had given to the agent saying, you know what, you can access all this structured data and all of this unstructured data. And the agent, the context layer of the harness, figured out that these were the memory or the conversations that people were having with the various agents. And it's not just the raw memory, it actually creates learnings or facts from that memory. The fact that the customer prefers to have fulfillment before promising revenue, or sustainability is a non-waivable line. Those are not in the actual CRM thing. It's in some unstructured corpus somewhere or in some conversations that the agent had. All that becomes memory and part of the context. And Of course, you wanna interpret some of the semantics of what it means. What does it mean to do availability promise, ATP, and so on and so forth, along with all the governance rules? And that's how this customer commitment agent got the right dynamic context. Not all the data, just the subset of the data needed for it to make that happen. Now, we also remember we had the same thing about it shows Northwind Traders rather than all the other 3 customers. How did that happen? Same thing. Well, we had all of this context back and it was able to go figure out that from one of the conversations, the deal condition was a 20,000-unit commitment across Europe for it, for the deal to go happen. It was able to pull all that stuff from different places. And again, interpreting what Q3 meant was really October, not September, along with the governance thing. And it was able to zoom in on this particular customer for us. for Maya to focus on. So that's how you can bring in data, create context, and build agents, whether on Salesforce or non-Salesforce. Now, once you have done that, the first question your IT team is going to ask, okay, how are you governing, governing this? Are you making sure your agents are getting the right data? What about data leaks? Or how are you governing the agents? We have data governance and AI governance built right in. So here, as you can see, as the data is coming in into the context layers, automatically we tag everything. We put PII tags and all the other classifications, data classifications. And now you can look in a glance at the quality of the data, what's happening, and you can create sophisticated data policies. Like I can create a masking policy right here that says, you know what, I don't want the LLMs to see all my supplier data and others. You can literally just go mask it. So anytime an agent is asking that question or a human asking, these masks and data governance policies will get applied. And same thing for the AI governance too. You can apply all of the policies in terms of compliance, observability, all of that on all of your AI agents as well. So governance at both the data layer and the AI layer is built out of the box. And of course, you want to observe all these agents, how they're doing. You can have full monitoring and evals. Eval is a new way of doing quality control of agents. So we let you create AI-based evals for any agent in your enterprise, and you will be able to go make sure these agents are working as predicted. And also, if you make changes to your agents, it can make sure whether it's still working the same as before. So that's all built out of the box. The other big thing is security. We saw how when Maya was trying to update that particular margin exception using the cloud agent, the cloud agent could not update it.

Why?

Because we have this new thing called an agentic identity that we have added to the platform. Basically, in this case, you have Maya, the person, and then you have the agent, which is acting on behalf of Maya. And you can now actually have roles and privileges for users as well as agents. You have full— exactly like you would treat a user, you can treat an agent and add permission sets and so on. So in this case, if I go back to the commitment agent, that revenue pricing exception, if you remember, Maya could update it, but the agent could not, which is why it was denied. But she could do the personal scope to say, you know what, I approve this particular update, and then the system would automatically do it. That is agent security for you built right out of the box. And last but not the least, we all worry about how much the token burn happens, and we have cost management for all of these agents because all these agents are going to go through a single gateway that we can control. So that means you can see which models are costing what. You can also create these model wallets. So you can say, you know what, I'm gonna give $100,000 for this model or $50,000 for this model and so on. And we can do semantic routing automatically along with even suggesting what those potential optimizations are. So that is what the power of the Enterprise AI Harness is all about. If you switch to the slide, Basically, what you saw was the ability for you to bring all your data, create the context, be able to go create agents, us or other agents, link them all together, see the governance in action at the data and the AI level, and be able to do security as well as cost optimization. Now, question you might be having is like, oh my God, this is like a visionary demo. How long is it going to take? Guess what? You already have all of these things available today. With Informatica, you can do all of that data cleansing, getting the data ready. With Data360, we just announced the ACE engine. In addition to CDP, you can now get all the context ready for your agents. With Agent Builder, of course, you have the ability to create all these agents. With Agent Fabric, you can have the full AI control plane today that can manage all of your agents across your enterprise along with all the cost optimizations. And the newly released Guardian that where we have the full agentic security, can give the power of all of this. Of course, Tableau is always there to give you that business semantics of what it means to do all of this stuff. So that is the power of the Enterprise AI Harness that's going to bring all of these technologies together and make the AI agents work for you, for your business with the right trust and governance. Back to you, Rohan.

Thank you, MK.

Wonderful.

Thank you, Gabe. Thank you, MK. You know, hopefully, you know, I know it was a long but a super comprehensive demo of how the different capabilities of the harness come together. This, like I mentioned, is the future of the Salesforce platform. We are relying on a lot of the capabilities that we have today, but really, really excited about where this is going. Now, please welcome Jason Lemkin, the CEO of SaaStr, to come and join me for another customer story. Jason is one of the leaders I'm super impressed by who's a CEO, can do all the business stuff, but is very detailed in the technical understanding. So I thought it would be a great experience to talk to you, Jason. So now you've seen the harness.

Yes.

I know you speak to a lot of leaders, CEOs. What do you think? Like, is this whole approach that we're taking, is it making sense?

It's what you need. So we've been running our little 8-figure business with 3 humans and 21 agents on Salesforce for over a year. We've been headless on Salesforce for over a year. So early. And the stress isn't the agents. Agents are great, right? But you can't trust them. You need a harness. And so we picked Salesforce over a year ago. And this is what we've seen today is much more powerful than what you could do a year ago. But this is what you need. This is what you need because you need a way to trust these powerful agents.

Thank you, Jason. And, you know, like, as, as you know, the models, of course, are very powerful, but models are going to be available to every enterprise. So as you think about your business and the others, How are you thinking about where the moat creation is happening and how does that get transformed?

You mean what's more valuable, the model or the application or the harness?

That's right.

Let me step back. If you're actually building yourself, building your own agent, building your own application, model is the most important. We use Fable 5.1, which is the latest from Anthropic. You should use it.

It's badass.

It's actually much better than the models a couple of months ago. And the models got much better if you're building your own. You should not build your own though. You should buy. You should only build what you can't buy. And so everything that you're going to buy, or more importantly, everything you want to do with your data, everything you want to manage with your data, then the harness will figure it out. The harness will figure— and just like we saw today with your CTO, the harness will figure out which is the right model. The harness will use the router and figure out which one is the best for different workflows. You should not have to worry about this stuff anymore. The harness does. And we get confused on social media. Well, do I have to pick the model? Do I have to pick the— it's confusing. Yes, Fable 5.1 is great. But for 99% of what you're doing, you don't need to think about it. Just have the harness figure it out for you. This is 2026.

All right. No, thank you, Jason. I think that's actually wonderful guidance, which is if you had something like the Enterprise AI Harness, your point essentially is use it. Use the intelligence because it's going to do the right thing for you. Use it.

Yeah. Use it.

Well, thank you so much, Jason. Yeah, thank you.

Thanks for everything.

Well, again, hopefully this gave you a sneak peek. Of course, I highly recommend you go look at a lot of the other sessions that go deeper into the platform products that we have. MK mentioned Informatica, Tableau, you look at MuleSoft, our Data360, the core platform, etc. These deep dive sessions are going to give you a lot more information, but if there's one thing you walk away with, is models will change, agents will change, you'll build new agents, but your IP, your customer data, your enterprise context is what's gonna make your company unique. And our belief with this future of the platform, which is Enterprise AI Harness, is we are gonna help you create it, govern it, and secure it. So thank you and welcome to Dreamforce. Thank you.