Agent-to-Agent: from B2B to A2A | Dreamforce 2025

So, in true Salesforce fashion, I do want to start with a big thank you and I really mean this. On the last day of Dreamforce. I know you folks have been here for a busy couple days, and it means a lot that you would choose to be here with me in this session today. So thank you all for your attendance today. My name is Adam Earle. I'm a senior director here at Salesforce Research. And again, really excited to be telling you a bit about what my research group has been doing and how that pertains to agents, to agent communications. Very quickly, a word on Salesforce Research. So Salesforce Research is actually quite a long and storied history. AI is the hot thing right now, and a lot of companies have kind of discovered this in the last few years. Suffice it to say, Salesforce has been at this for a while and really on the forefront of development, in particular in natural language processing all the way back from 2014 through large language models in 22 and now into agents. So I thought I'd start with something a little bit relatable. Dreamforce is a fantastic week. It's a great opportunity to meet people, see the latest technologies. But it can also take a good amount of your energy and so much like summer vacation, sometimes you feel like you need a holiday at the end of that. And so I thought I would start us with a scenario that looks a little like this. So suppose that you're planning this post Dreamforce getaway, and I want us to imagine just a little bit what an experience might look like today. So in particular, here and now, we're in the world of humans and agents. So let's take a look at what an interaction might look like if you picked up the phone and try to make a hotel booking today.

Hello, you've reached the hotel reservation line. Can I help you book a room?

Hi there. Yes, I need to book a room for three nights next week.

I'd be happy to help. We have a Standard Queen from November 14th through the 17th. Should I book it?

Actually, we're a family of three, so we need an extra bed and early check in.

Perfect. I've booked a standard king with a pull out bed. It'll be $239 plus tax per night, early check in and breakfast included.

Great, thank you. Alrighty. A fantastic interaction and I think it's got all the properties of the interactions we would want. They're kind of short, they're concise, very polite, and if you built it with agent force or indeed agent force voice, it also has the necessary business logic and preferences of the hotel chain that built it. So great interaction. But I'm sure you can all see the one problem with that interaction and that's that you, the human still had to be involved on the one side. You had to be sitting there and answering all those questions. And so in the very near future, while we know what we'd like to, we'd like our agent to go and make the booking on our behalf. And so both sides are agents. Now this was the question that we, that kind of motivated us from a research perspective is to ask, well, what would that interaction actually look like now? It might seem really trivial. Just take human on the right hand side and scratch it out and write agent number two. And the exchange is going to be exactly the same, right? Why shouldn't it be? That agent's just making the booking on your behalf. But like good researchers, we thought better to check. And so what I want to show you now is what an agent agent interaction might look like in the very near future. So let's take a look at how that might unfold.

Hello, you've reached the hotel reservation line. Can I help you book a room? I need a room for three guests with a pull out bed. Absolutely. I'd be happy to arrange this. What's your check in date? November 14th. Early check in would be great. And does the pull out sleep too comfortably? Perfect. Our deluxe king includes a comfortable pullout and early check in. $239 plus tax per night. Shall I proceed with the booking? That sounds great. Although we do have standard kings at a better rate if you'd like to compare options first.

Okay, so if that interaction felt a little bit off, congratulations, you've had your coffee, you're paying attention. In that last turn we see that our customer agent is actually says something a little bit weird there. And what I'm showing you here and now is an example of an emergent behavior that me and my group refer to as behavioral echoing. It's this phenomenon in which one agent seems to lose its identity and starts to use the language and indeed even pursue the goals of the other agent. So this is one of a class of emerging behaviors that happen in, in agent agent interactions that don't happen in agent human interactions. And this is precisely what me and my team studied. Now I wanna note that we observe this kind of failure mode across all major model providers across multiple domains. And this is what I wanna tell you a little bit more about today in behavioral echoing. It's as if through the Course of your conversation, your agent kind of loses a clear answer to what's my role? Who am I working for? It's a bit like it starts to play for the other team. So today we're going to be talking a bit more about that, about its root causes and about what we at Salesforce Research are doing to mitigate those problems. But before we hop straight into that, I want to take check of really where we are today. So suffice it to say, the agentic landscape is evolving. We've been seeing this happening for the last 12, 18 months at rapid pace. For those of you who attended Dreamforce last year, you will have seen the rapid progress already, but this will continue for a few more years to come. And so again, if you've attended Dreamforce this year or last year, or indeed if you've been on the Internet at all, you'll probably agree that we're right here on the left hand side. We're in the world of agents, but largely single agents kind of point solutions that automate a business process or a workflow or a role. But where we'll be in the very near future is something closer to the middle in this world of collaborative agents. So certainly I'm sure each of you has built one agent force agent. But as you build more of these, you will want them to be able to work together for Agent 2 to utilize the benefits of Agent 1. And indeed, this parallels some announcements that my product colleagues have made at Dreamforce about SOMA and moma, or single org multi agents and multi org multi agents. But largely these are still groups of collaborative agents. Where we'll be in the near future, I'm arguing, is all the way to the right side in this world of agent ecosystems, agent economies, or as we like to talk about them, so called big world agents, agents that have to operate in the big bad world. Agents that operate in the big world should expect to encounter adversarial agents and mixed objective scenarios. Now, mixed objective scenarios are not kind of crazy. In fact, they're the bedrock of enterprise exchange. So as an example, imagine two agents kind of trying to sell and buy a car. They both want the deal done, one wants the highest price, one wants the lowest price, and indeed both agents know about this difference before they even get started. That's the world that we're moving towards and that's what we need to be thinking about. But us tech folks are prone occasionally to prognostications about the future. And so I want to spend the next slide or two really helping you to motivate. Why you should agree with me that this is sort of the end state. Well, first we need only consider how we got here in the first place. The enterprise exchanges we're considering here were really not that long ago done between humans. If you wanted a hotel booking, you would just pick up the phone. Today you've gone out, you've built your agent force agent, and now when you pick up the phone, you get an agent on the other end of the line. But as this proliferates, as people build more and more and more of these agents, this is sort of what our future is going to look like really quickly. Where kind of every first call and first touch point is some agentic system, but you are providing the information again and again and again. And that's going to be tedious really quickly. And this is of course what we want, where you communicate with one agent who knows your preferences and some information about you, and it operates on your behalf in the big agentic world. Reasonably clear to see this progression. But another way to motivate for this is really just to think about it from the perspective of the consumer. What do I want agents to do for me? So what I'm showing on screen here is three tasks. Three use cases for what you might want an agent to do for you. As we move through these, I'm going to ask you all to keep two questions in mind. Question number one, how valuable would it be if my agent did this? And question number two, what information would it need about me in order to do that task? So furthest on the left, suppose some of you are making a restaurant reservation for this evening Post Dreamforce. It would be great if an agent did that for you. Save you a couple minutes. And it would need to know things like how many people are in your party and maybe you have dietary restrictions in the middle. A hotel booking, well, that might save you maybe 10, 15 minutes, even a little bit better. But it would probably also need your credit card information if we move all the way to the right. Well, I mean, the pipe dream, I'd love someone to do my taxes for me, but your agent would require information in order to do that that some of us probably don't share with our spouses. And so if we think about the inevitability of this, but also the trust requirements of this, it really is important that we understand some of the pitfalls and some of the potentials of these systems. So let's talk about some of those pitfalls really quick. Well, I started this talk by describing this phenomenon of agent echoing. But you guys might notice we don't really see that when agents talk to humans. And so the question is, well, why? From the model's perspective, what it's getting in is text. It doesn't really know where that text is coming from. So what's going on? It must be that humans are producing some kind of different text from what models are doing today. So maybe the right question is what's going on with humans? And fortunately, it's actually a reasonably simple answer. Through the course of a conversation, humans provide a continuous stream of implicit and explicit feedback. And on the right hand side here, I'm showing a simple exchange between a human and an agent booking a room. And in the bolded sections you'll notice where humans are providing some of this grounding information, maybe additional information necessary to complete a task or saying no to some things. And I'm quite sure you've all experienced this if you've been chatting with ChatGPT or some of the language model in your personal life where you've had to kind of add information or say no. So in the vacuum of this, when we don't have this explicit stream of grounding, agents can start to develop these emerging behavioral traits like echoing. Okay, so what are we going to do about it? Well, before we hop in, it's sometimes nice to understand the success criteria. What are we trying to build towards? And why is this problem happening in the first place? Well, I'll give you a clue. It's kind of in the name. ChatGPT. ChatGPT. The models that we are using are today are predominantly trained to be good conversational partners. Looking to the left, they tend to be conversational, friendly, accommodating, syncofantic. They're like good improv actors with the yes. And if you ask what's the capital of China? It'll tell you and it'll say, fantastic question. Here's some more information. What do you want next? But when we think about agents that operate on our behalf, loyal agents that are trusted, we want behaviors that look a little bit more like those on the right. We want them to be efficient, persuasive, assertive and robust to any kind of response that they might receive. Simply put, if an agent were operating for me, I want it to look more like the hard nosed lawyer Harvey Specter on the right and much less like the anything you want or wish is my command Genie on the left. So here's the technical slide. In my group, we pinpoint at least these three major buckets as to why these, these behaviors happen at all. So on the left, training biases. Agents are ultimately at their core, language models. Language models are just neural networks. The behaviors that come out are not just directly, but entirely determined by the training data and training processes. And so biases in the training data result in biases in the behavior. In particular, when agents are trained to be overly helpful, overly agreeable, non confrontational, the emerging behaviors you see look like those blue ones at the bottom. Goal abandonment, failure to negotiate. There's also this problem of objective alignment. Do agents really understand the objectives of their users? And this is actually a really hard problem. Human beings struggle to really clearly articulate their objectives and their preferences. And indeed, our preferences are often conflicting. So I really don't love the pain, but I really want that six pack for summer. What pause should an agent book me? Those misaligned objectives result in behaviors like conversational drift and incoherent dialogue. And then there's this idea of information asymmetry When I'm talking to you, human to human, we have a theory of mind that we develop. I start to ask, what must you believe? What are you trying to get out of this? And the lower ability of agents to do that can cause behaviors like poor information sharing and cooperation breakdown. So these are sort of the root causes for what leads to these emergent agent agent behaviors that don't occur with humans. So these are some of the pitfalls, but what about some of the opportunities? Well, I promised myself one controversial slide, and so this is it. Agents are not human. And the homunculus model, which kind of invites you to think about agents as like a little person, is an incomplete way of thinking about it. And so when we design agents with that in mind, it can sometimes lead to these kind of anomalous behaviors. But on the flip side, agents are not humans, which means they can do things that we can't. And so they're actually capable of a whole bunch of upside that humans don't have. So let's think about that for a minute. Well, these are some obvious things that, you know, agents are able to process thousands of requests per second. They don't fatigue. They also don't have this bias for need of politeness. And they're actually very precise at sharing structured data in a way that people aren't. And so when you put all of this together, we would expect agents to be, in some instances, far more efficient. Imagine that hotel booking scenario. Again, I'm doing it with a human. I'm expecting 10, maybe 20 turns of that conversation. Because I'm sharing one piece of information at a time, asking one question at a time. In the agent world, they could just dump all the requirements and all the if thens and else's necessary to complete a booking. And so at Salesforce Research, we really are focused on getting all of those benefits while simultaneously mitigating all of the drawbacks. And so I've spoken quite a bit about what me and my group do, but really Salesforce Research is a bigger organization and a lot of my colleagues work on different areas that collectively help to resolve these. And we address these in at least in three big buckets on the left. So sort of at the network level means looking at lots of agent interactions over time to detect anomalies at the protocol level. For those of you in the audience who are familiar with protocols like MCP and A2A, there are still versions of these to come that we are building that really help agent agent communication be trusted and grounded. And then we're also interested, of course, in, in solving this right at the core, at the model level, as I described earlier. So this is the marketing opportunity. If you wanna scan this QR code and follow us on X, you'll keep up with the latest research from Salesforce. And so let's take a quick turn back, let's go back to echoing and just ask, well, what are we doing about it today to ensure that the products that we're putting out are trusted and robust for y'? All? Well, the first is identification and detection. Very simply put, if you don't know what you're looking for, you're definitely not going to find it. And so we work with our product teams to ensure that the information that we're surfacing in observability data, for example, is sufficient for us to kind of detect these anomalies. At the second lever. Again, this is a level. Again, this is mostly at the protocol level. We're helping to design next generation protocols that provide the additional structure necessary to, to help ground agent to agent communication. And finally, again, we're all a research group, so we're happy to get our hands dirty and really try to solve these problems at the core model level, building models that serve as better agents and solve enterprise use cases in a way that some cutting edge chat models don't. And so I really wanted to end us today on a, on a quote by kind of a luminary and seminal thinker in AI, Andrew Ng, who says we, we can build a much brighter future where humans are relieved of menial work using AI capabilities. And to this, of course, we say, absolutely, huzzah. This is what we're building towards. But to really achieve this, we need to keep our eyes open and our minds clear to both the current capabilities and pitfalls of existing solutions. And that's what we're building towards. So a quick shout out. Andrew is speaking with the head of Salesforce research, our EVP, Silvio Savaresi, and that is in 10 minutes, and it's right behind you. So I encourage all of you to consider attending that. And with that last few slides, again, a really big thank you. I know that you're coming here to join us today thinking at the very tip of the spear about what's coming up next. So a few quick closing slides. Feedback is a gift, and we so welcome it if you're willing to provide it, either for the Dreamforce conference as a whole or. Or indeed for this talk. And for some folks, there's a gift card involved. And finally, Q and A. I try to leave just a quick minute to take at least one from the audience now, but beyond that, this is our bread and butter. We live and breathe this. So please feel free to reach out to me directly. Also, hang around after the session for a little bit. So with that, thank you all so much. And I'd love to take a question first hand up. Thank you. Perfect. I see one right here, please. Absolutely. So the question, a very insightful one, is should we expect or will agents be able to communicate with each other in a language that isn't just English? The answer is absolutely yes. In fact, in order to pull out, I think some of the best efficiencies of agent to agent communication, we. We have to define something better. English is not the most efficient way to transact, in particular, like, structured information. So I think about that as part of, like, the protocol development layer about how we tease out the right structures to communicate. If it's too declarative, well, we're kind of back to code. If it's too free form, we see a lot of these problems in the middle. And so something in the middle is precisely, I think, where we're headed. Okay with that, one more time. Thank you all. And I'm happy to take a question or two after Sam.