Use Words, Not Code with Agent Builder | Dreamforce 2025

As is customary with any Salesforce presentation, we want to begin by saying thank you. Thank you so much for spending your time with us, for coming out to Dreamforce, making the time investment, the financial investment. We know that it's a significant commitment to come out here for a week. And also, on a more personal note, thanks for coming to this session. We know it's at the end of the day, but you made the right choice coming to ours. There are hundreds of other sessions you could have chosen, and you chose ours. We really appreciate that. So thank you very much. We won't let you down very quickly.

My name's John.

This is my colleague, Evan. We are both designers on the Agent Force team specifically focused on Agent Builder, and we are really excited to share with you some of the new and exciting features of Agent Builder. So we all know by now we're probably a little bit fatigued by AI, AI, AI. But we know that there's a ton of potential in it, and it can do really, really amazing and magical things. It can generate new ideas for us and new content. It can analyze massive amounts of data for us. It can even transcribe sales calls and pull out next steps for us to make sure that we're doing the right thing and having the right conversations. But it's really our responsibility as designers to make sure that we are channeling all of that potential so that the agents that we build and the ways that we use it, it's all for the right reasons. And the agent is doing things that we expect it to do and. And not doing the things that we don't want it to do. And so we're excited to tell you more about that today. But first, Evan's gonna tell you a little bit about why building agents is challenging.

So agents really have to do two things in order for them to be useful. They have to understand people and understand how they speak, what they want, what they need. But they also have to act with the precision and execution of software. And this is what makes agents incredibly powerful, but also a little bit difficult to design. And when it comes to how an agent works at every step, it's not just a single step. It's really a chain of responsibilities. They have to interpret natural language in the way that humans actually speak it. They have to recognize user intent, not just match keywords. They have to choose and execute the right actions every turn, every time. And they have to present those results back appropriately. Not just appropriate to the end user and how they would understand it, but also appropriate to the modality or the surface that the user is using in order to interact with the agent. And what makes LLMs so powerful is also a little bit of a double edged sword. They're incredibly flexible. They can understand that human input. They can reason through a large list of instructions, they can determine what the next best step is and then they can actually execute on that. But all that flexibility also introduces variability. And sometimes you don't want variability when it comes to designing your agents. Variability sometimes can get in the way of that trust layer. Imagine a healthcare use case, for example, if your agent is providing discharge instructions for a patient. You don't want those discharge instructions to be variable at all. You want them to be followed like the letter of the law, every time, at every turn. And it's safe to say our expectations are shifting. I'm sure many of you in this room have used LLMs in your professional lives. I'm sure you're using them almost every day, maybe even today. And natural language is becoming that interface that customers expect. We're getting more and more comfortable getting work done using nothing but words. And we want to just type what we want and then see it come to life. And AI is lowering the barriers and broadening who can build. This means more people and more roles with less technical ability can build powerful software solutions without even learning how to write a single line of code. And our expectations for the experience are generally higher. Tools must feel simple and fast and accessible. And anything less than that can feel immediately dated when we compare it to the tools that we're using in our personal and professional lives.

So expectations are changing. So what does that mean for the new agent Builder? Well, for us it means a brand new, entirely new way to build. It means assistive and creation so that you're not doing it by yourself. It means reliable when it's out working, when those agents are out there doing things on your behalf, it's doing it consistently and reliably every single time. And of course, finally, it means it's transparent. So that if you are troubleshooting and you're trying to track down things that maybe aren't working the way that you would expect, you can do that easily and you have the tools to do that. So let's take a look at what we're calling building agents in your own words. Now, we're going to go through a couple different examples of some of the new features in the new Agent Builder. They're videos, but they're short, they're going to be looped, we're going to talk over them. And so if you miss something the first time, don't worry, it's going to come right back around. For the sake of our example, we're going to be building an agent for our bed and breakfast called Cozy Haven. And we want to make a concierge so that our visitors have the best experience possible. If they want to order breakfast, they can order breakfast. If they want to make changes to their itinerary or see what local attractions might be, they can do all those kinds of things. So the best place to start is to get started and create that thing from scratch. And so what better way to create, we think, from a design standpoint than simply describing what you want? So here on the screen you can see that I'm simply just describing that I want to build an agent for my bed and breakfast called Cozy Haven. I could have started from a template. We have lots of great templates and that's a good way to start too. But wouldn't it be great if I could just describe the kind of agent that I want? What it's doing is it's actually looking at my intent, it's breaking down the requirements that I've given it and it's transforming all that into real agent script. If you saw maybe some of the AgentForce keynote earlier today, it's you'll be familiar with AgentScript. If not, you'll see a little bit of that later. And now we've ended up inside of the agent builder itself where it's done a ton of work for me already. Like I said, don't worry, we'll come back around. It's created topics for us, it has configured important settings for that agent. And now I'm ready to continue working, to continue iterating and to continue testing all along the way. So now that we've built our agent, of course we need to iterate. And just like creating from scratch, it would be really nice if we could iterate and refine conversationally. And now you can. With the Agent Force Assistant, you can see on the right hand side, it is just as simple as telling it exactly what you want to do. So let's say I want to add some new instructions for my agent around the hours of operation for room service. And our hours of operation happen to be 6am to 10:30. And we want to make sure that if someone who was up late and wakes up at 11am and wants to order waffles and pancakes and things like that, that the agent will very generously say, unfortunately, our room service for breakfast is closed. At this time, but we're excited to serve you for lunch. So that's all I did here. I simply just asked the Agent Force Assistant to add some instructions, gave it the hours of operation that we wanted to add. And, and what it's gonna do is it's not just gonna look at the dining and breakfast topic in the instructions here. It's actually looking at the entire agent metadata across all the topics, all the actions and configuration settings, and it's looking for all the different places that it may need to make modifications for our new room service hours. Now, in this case, this is a little bit more of a simple example. It just made changes here in this topic, but if there were multiple changes, you would've actually seen those reflected across all of those different places. And then I could paginate through them and I could see which ones I want to accept, which ones I may need to further refine, or maybe which ones weren't quite right for me. But I think this looks great. So we'll go ahead and accept it now. That's great, because sometimes I want to add instructions, but I'm not exactly sure where to add them. And so that's where the Agent Force Assistant is really powerful. But sometimes I do know exactly where I want to make changes. And it would be really great if. If I could just use AI right on that specific block. That's what we call these inside of this canvas here, we call them blocks. So in this case, on the instructions, I want to make sure that if a guest is a vip, that they get priority service. I want to make sure that their waffles make it to their door in five minutes or less. And so traditionally, what you would do is you would write an instruction that says, if the guest is a vip, then make sure that they get priority service. But we know that LLMs don't quite work that way. They don't always follow the instructions that we give to them because that determinism isn't sort of built in. Evan's going to go into that a little bit later, but that's what I'm doing here. So I'm just going to highlight that block and I'm going to say I want to make sure that these instructions are followed reliably every single time. So I'll go ahead and submit it. The agent's going to think, and it's going to make some changes for me. I can quickly review those. If everything looks good, I'll go ahead and accept those. It really is incredibly simple to do.

Yeah. And A lot of tools can generate the right outcome once, but the real challenge is making sure that that same outcome is generated every time that the agent is run. And we've learned a lot about how our users are writing their agent instructions. A lot of the times you're using the word always or never, oftentimes those words are in all caps, trying to kind of bully the LLM into doing what you want it to do every time. But that is why you want that extra level of control on your agent, and that is why that we have built more deterministic controls that you can inject directly into your agent instructions. So this hybrid reasoning allows for introducing if statements or as action calls directly inside of your agent instructions. Make sure that you produce consistent, repeatable outcomes every time the agent is run. And if something doesn't go as expected, let's say you're simulating your agent. You have AgentForce there to help you every step of the way, even when you're debugging. So I'm simulating this agent and I've asked it a question and it's generating an answer. I can see the trace of every step that it has taken when it's generating that answer. But if something doesn't go right, I can click that Ask Assistant button and not only will it help you diagnose what went wrong during that turn, and if you want to add more context to the question to help debug, you can. But it'll also offer up suggestions on what you could do to fix it, and it'll even go as far as fixing it for you. So we've accepted those changes and it's go ahead. It went ahead and made those updates inside the canvas, which will also cascade to the agent script. So you can spend less of your time combing through error logs and finding out what went wrong, and more time building the right agent. And every step of the way, behind the scenes, agent script is being generated on all of your instructions.

So.

So whether you started in the AgentForce panel on the right and started to build conversationally, or you started writing natural language instructions in the canvas view with your conditional logic and your if statements, you can always toggle over to the script view to see that deterministic agent metadata being written alongside every step. So we wanna meet our users where they are at, whatever technical ability they have. If you are familiar or comfortable writing in a scripting language, you can just go right over to the script view and start writing your agent metadata there. Or if you're more of a business user and you want to build conversationally with AgentForce, you can. So there's a starting point for really everyone.

So those were some quick design hors d'.

Oeuvres.

If I can use that right before happy hour and dinner. We know it was maybe a bit of a fire hose of new features, but we're super excited about them and we think that you're gonna love them. Quick summary of what we talked about, starting from natural language. It's easy. It's just like talking to a colleague and describing the type of thing that you wanna build now you can do that. In the new Agent Builder. We also saw how once that agent has gotten going and you're starting to make refinements and iterations on it, you can do that conversationally too. In the new AgentForce assistant panel, we looked at inline AI assistants for times when maybe you know exactly where you wanna make changes and you don't want it to affect the broader agent. And Evan walked you through debugging. We have an amazing new debugging area where you can see exactly how your agent is thinking under the hood. And if you want to dig even deeper into the detailed execution trace in that debugging drawer that comes up, we have that too. Lastly, we looked at deterministic control to make sure that your agent is always doing the things that you want it to do and never doing the things that you don't want it to do. And then finally it is code ready. So whether you are a business user who prefers working in maybe more of a document style editor, or you're more of a technical developer who prefers working in code, we serve both types of users. We really think you're gonna love it. Thank you so much for coming to our session. We promised you we would get you out of here in time to make it to your happy hours. Thank you again for coming. But we do have one more ask for you. If you would all get your phones out and hold them up. Get your phones out. Get your phones out. I can kind of see you get your phones out. Amazing. If you could scan this code for us and take this survey, it would mean a lot. It's really helpful for us to know what went well, what we can do better, just to make sure that future dream forces are just as amazing as this one. And with that, we are so excited for the new Agent Builder. We hope you are too. We hope you have a wonderful rest of your day. Thanks for coming out.

Thanks.