Christophe Coenraets, SVP of Developer Relations at Salesforce, tells Eira and Ben about building the new Salesforce Developer Edition, which includes access to the company’s agentic AI platform, Agentforce. Christophe explains how they solicited and incorporated feedback from the developer community in building the developer edition, what types of AI agents people are building, and the critical importance of guardrails and prompt engineering.
The new Salesforce Developer Edition includes access to Data Cloud and Agentforce, Salesforce’s platform for building, customizing, and deploying autonomous AI agents. Developers can sign up here to start building.
Connect with Christophe on LinkedIn.
Instead of a badge, we have a question for you: If you are a developer who is working with AI agents, what are you building? What tools are you looking for? What problems are you interested in solving with AI tools? Send us an email at podcast@stackoverflow.com.
[intro music plays]
Eira May: Welcome to the Stack Overflow podcast. My name is Eira May. I am the Business Content Editor at Stack Overflow, and I am here today with my old friend, Ben Popper, co-host of the podcast.
Ben Popper: Hi, how are you doing?
EM: It's good. Good to be back in the same recording studio with you, Ben.
BP: Congratulations on your new title. ‘Business’. When I think Eira, I think business.
EM: That's my jam. That B2B stuff, that's my jam.
BP: Exactly. So you went to an event recently and met some cool folks. Agentic AI is a word I can't get away from on my social feeds and in my news feeds. Tell me a little bit about who you met, who we're gonna be talking with today.
EM: That's right. Yeah. So I went to Salesforce's developer conference in San Francisco in March, and one of the folks that I talked to about Agentic AI and the kind of Agentic AI developer tools that are in the works was Christophe Kratz, who is the SVP of Dev Rel over at Salesforce. Thank you for joining the show.
Christophe Coenraets: Thank you for having me. Great to be here.
EM: Yeah. So glad you could reprise some of our conversation that we started back at TDX. I wanted to ask you first off, just what was your path like to getting into the work that you do now, to the role that you have now at Salesforce? How did you get started working in software and technology?
CC: It was a very, very, very long time ago. I won't even tell you the language that I started with because that would age me. But yeah, I was a developer myself for a very long time. I definitely did the Java, J two E path and then different languages. And then I kind of found out about developer evangelism and it was really, it became a passion for me because it was the intersection of my passion for software and then a passion to kind of inspire people and kind of enable people. And so, I've been in developer relations ever since, probably now for 20 plus years, and I love it. It's a great job.
BP: I know that you're, when you were at this event, kind of focused on a developer edition of Agentforce. My only interaction so far with AI and Salesforce is that they have an amazing commercial with Matthew McConaughey. I know it gets left out in the rain and it has something to do with people not correctly programming reservations at a restaurant. So maybe that's a good place to start. Christophe, what is a developer edition of Agentforce?
CC: You know, first of all, the developer edition has existed for a very long time at Salesforce, and the basic idea was to allow developers to experience the platform and really, you know, start developing without necessarily working for a company that uses Salesforce. So, it's simply an environment where you can explore the technology, you have pretty much full access to the technology. Now, you know recently, obviously we released Agentforce and that technology along with Data Cloud that we can talk about in a moment. But that technology was not part of the developer edition, mostly because it's a little bit of a different model, it's a consumption based model. You know, you make requests to these large language models and it's slightly different. And the feedback that we got from the community is, you know, we love what you're saying, we see you at every event, you talk about Agentforce. But as developers, the only way we will know if it's awesome is if we can play with it, and that's what developers do. And so what we wanted to do was to give them an environment where they can experience Agentic AI, where they can build agents and kind of feel it, you know, think about use cases. So, you know, in a nutshell, that's what the Agentforce Developer Edition is.
EM: Could you just give us an overview kind of what Agentforce is? And where the data cloud piece comes into it, in terms of the developer edition, and what that changes?
CC: Yes, definitely. And you know, I'll speak about it from a developer point of view because there is Agentforce, the environment that, you know, a business will experience and then we define it in a certain way, but I know here that the audience is developers. So from that angle, Agentforce is really a platform that lets you build agents and these agents can, you know, obviously do all sorts of things. They can work in an autonomous way. They can also escalate to a human, you know, if that's what is needed for a specific task. But at a super high level, Agentforce is that platform giving you not only the platform, kind of the run time if you want, but also all the tools that you need to actually build these agents, to test these agents, because that becomes very important as well. And, and also to monitor them, you know, what are they doing? Does that look right? Can we fine tune them? So, at a super high level, that's what Agentforce is.
BP: Maybe could you give us an example of the top two most popular types of agents that you see people building and that will give us a sense of how they work and you know, how developers are using the tools on your platform to create value.
CC: I can give you two categories that I think are really kind of important because we start to see two different types of agents that people are building. The first one is employee agents, so, you know, whatever business you work on or a specific customer works in, whatever industries, these are agents that as an employee you use to make your job easier. And so you can ask the agent all sorts of things that typically either you would have to navigate multiple systems to get an answer, or you wouldn't even know where to go to get an answer or to perform tasks. Because agents are not only about getting information right, and you know, like the first types of agents that you see, it's often about getting information, but I think the most fascinating part about agents is that they can also take actions, right? So imagine as an employee anything that you may want to do, you don't know how to do it, if you work in a large company, they may be like, you know, dozens if not hundreds of systems. Well, an agent gives you that conversational interface where you can simply say what you want to do. The agent will figure out how to do it right. So that's employee agents. The other side, the other type of agent would be more of a customer agent, customer service type of agent. It's essentially the same thing. How does your customer interact with your company? Well, they can use the phone, they can use your website, but often times they don't know how to get information. They don't know how to do certain things. So, these agents can really help them 24 hours a day. And that's part of the magic here, is that these agents are available nonstop. So employees, customers, customer service, we see a lot of that.
EM: One of the kind of metaphors that I've been kicking around in my head since TDX is that if you think of a generative AI system as like generating a map for you to get from point A to point B, it seems like what an AI agent is really able to do, on an Agentic AI system, is to just like put you in the car and drive you there. Do you think that that's a sort of a reasonable metaphor? How far off am, I guess, in my thinking?
CC: I mean that that would definitely be a type of agent. I think to me, what's really fascinating, and that gets to how it changes the world of developers, is that in the example that you gave it seems like a pretty defined task, right? You're going to enter your location, or it's going to figure out your location and then you're going to say where you want to go. It's a very defined task, and in many ways we didn't need agents to do that. But somehow you could use an agent. As a developer, what you are going to start doing is not so much build these end-to-end workflows. Until recently if you had to build an application, you know, you kind of gathered requirements and you kind of figured out all the things that the application would have to do, and then you would get to work and start coding them. I think the major difference, if you think about agents, is that as a developer you don't always do that anymore. You start building a number of actions if you want, and these actions are kind of building blocks, and these agents can then orchestrate these building blocks in many different ways, including in ways that you had not planned before. So in a way, an agent can do things that you really never anticipated. And to me that's really what's really transformational about agents because I don't think we experienced that ever before as developers.
BP: Yeah. You know, one of the things that I feel like I've discussed over the last two or three years as this type of AI has made its way deeper into the technology stack, is that, to your point, it's non-deterministic. If you connect them up to a bunch of different systems and say that they can talk to a bunch of other agents in the company, you might ask for something and it might come back and do it in a way you didn't expect. Just like AlphaGo played a move we hadn't seen in 10,000 years of Go. Are there tools that folks can use to: a) to tell them what they can't do? That's probably pretty important. And then b) some kind of prompt engineering to give them the right tone, or the right language, or the right framework for like you said, going out and exploring how to get a task done.
CC: Absolutely. Because what I just described can look a little scary or not kind of realistic. If, like you say, it's non-deterministic and it can look a little bit, ‘okay, maybe it's going to do the right thing, maybe not’, right? Maybe it's going to start doing, you know, we have all heard these stories, right? Of agents getting out of control and starting to do things that they were not supposed to do. And so that is really super key. So it's really providing the guardrails to make sure that the agent stays on track, and you really define the parameters of, first of all, what an agent can do. And it can be as basic initially, as you define these guardrails, as ‘well answer that type of questions, but not those’. I mean, that's the very basic right, but you can go really deep in providing these guardrails. When you use Agentforce, you can do that using instructions where you know, you really can provide a set of instructions to make sure that the agent stays on track and really does what it's supposed to do, and doesn't kind of venture in territories that it's not supposed to venture in. So guardrails are really, really, really important. And then of course, you want a kind of a multi-layer approach. With Agentforce, we've been talking for a long time about the trust layer. So it's kind of a set of mechanisms to make sure that the agent is not going to hallucinate, that it's not going to engage in conversation that it's not supposed to engage in. You know, we have toxicity filters and all these things that make sure that the agent stays on track.
EM: I wonder if you could talk a little bit more about the prompt engineering piece there. I mean, like, we've heard a lot about the value of prompt engineering. When we're talking about generative AI systems, how does that sort of translate or change when we're talking about agents?
CC: Yeah, you know, in many ways. So first of all, you can use prompts as part of the process, right? So, if you think about these building blocks that I was talking about, right? How do you build them? Because if you think about it, LLMs can do a ton of things. Couple of things that they cannot do, by default, is to access your private data and then take action, which is really important here. So, how can you kind of, within an agent platform, like I described, how can you augment, if you want these LLMs with the ability to take actions, right? So you can provide actions in the form of an API call. You can provide an action in the form of a flow, a Salesforce flow, but it's essentially a workflow. You can provide an action as a piece of code, but a granular piece of code. But you can also provide an action as a prompt, right? Because part of the action that an agent may take, it would simply be the execution of a prompt. And so everything you learned about prompt engineering still applies here too. But there is another part that I think is also fascinating, we just spoke about guardrails and instructions. How do you instruct a prompt to do certain things? And not to do certain things? Well, oftentimes these are prompts too. How do you write these instructions in a way that are going to be really efficient? Right. And so one of the things that we realized is that some people are really good naturally at providing these instructions, in the same way they're good at writing prompts, but it's really almost like a science. And some of the things that we did recently that we added to the platform, with a new feature that we called AI Assist, is basically the ability to kind of help you have AI help you generate the guardrails, generate the instructions. You simply say what you want at a high level, which is what people typically do when they write a prompt, but obviously they don't provide enough information. And then AI Assist will help them formulate these guardrails or formulate these prompts in a way that's going to make sense and that's going to be efficient for the LLM.
BP: This is an interesting chain here, so I give a high level of what I want, and then an AI system helps me write a prompt, maybe with a couple of steps and some concise instructions and guardrails. That the AI who's receiving the prompt will be able to understand and articulate, in maybe a better way than the one I had write myself. It's like compilers on compilers here. You know, we're like -
CC: Exactly. Multiple layers of AI. Yes -
BP: Yeah, exactly. So you mentioned folks could write like a little bit of custom code. What's the reception been like from the Salesforce developers who are on the platform? What are some of the things they've asked for, or you know, things that you're excited to be able to roll out for them as you get their feedback?
CC: I mean, the feedback has been incredible. In the keynote at TDX, like a month ago, it was like big cheers when we announced the developer edition, you know, for a good reason. And I totally understand it, and I really pushed for it because there is that new technology out there that seems to really be super transformative. We talk about it a ton. People read about it and then they cannot experience it, right? So now anyone, you after we record this episode, you can go sign up for the developer edition. You can play with it and you can really feel what it's like. And so the response has been incredible. One thing that we did that super fun, you know, that's super fun at TDX we ran a hackathon and people got super excited about that. Now we'll have more TDXs, we’ll be in Tokyo at the end of the month, we’ll be in India in May. And we keep running these hackathons because the interesting thing, and when you ask me, what type of agents do people create? And I think to me the interesting thing now is the technology is incredible, the biggest roadblock that I see for massive adoption is that people have to start thinking differently. It's even bigger than that. But think about, when we went from text-based user interfaces to kind of graphical user interfaces, right? And people had no, okay, graphical menus, it was like, ‘how do I even use, why do I need’, it was a big transformation and people had to start thinking in a graphical way, when they had never thought about an application that way. And I think the same transformation, I think it's even way, way, way bigger, but people have to think about applications in a different way. What does it mean when the user interface is conversational? You know, and what are the use cases? That's also why this is exciting. The developer edition is exciting because it allows developers to really experiment and do crazy projects and try things. And then they will figure out some things make sense and some things don't make sense -
BP: Right -
CC: And I think we are really in a phase where developers need to experiment and figure out. Just like they figured out, throughout the many transformations of the industry, what makes sense and what doesn't make sense.
BP: Yeah, I think you make a good point. I remember Eira and I having discussions in 2023, everybody has seen that there's a new kind of AI intelligence that can be really powerful, but it's all tests and none of it's in production. And then 2024, a lot of this stuff is starting to go into production, and LLMs are getting to scale. And Agentic AI kind of feels like now, like you're saying, it's not that it's not in production, but people are still figuring out what to do with it, how to make sure that they can use it safely, experimenting with it. You mentioned, you know, they can work overnight, which is an interesting feature. You know, they can connect different data sources. I could see on the Salesforce website, an example where it's like a SDR. So you go to a big, you know, TDX event, you end up with 5,000 new email contacts and you say to the agent overnight, can you look through these? See if we have any of them. We touched on these before, if any of these are with companies that are in our target list or fit our ideal customer persona, and in the morning, give me a spreadsheet and if some of them are good, send them a nurture email. Is that the kind of thing, that's my platonic ideal of an agent, but is that something you'd actually build on your platform?
CC: Yeah. Especially the very last part of what you said because in a way feeding a system with a list of leads and then kind of somehow filtering them with some level of intelligence. You know, we've been doing that for a number of years. Again, what is really kind of transformative here is, like you said at the end, if that lead looks qualified, send them an email. It's not only sending the email, maybe you're getting a response to that email and it's like reasoning. Because that's also, when we speak about agentic, it's a lot about autonomous, it's a lot about the ability to reason, right? So now you send the email, maybe you get a response, well, what should I do with that response as an agent? Should I respond? Because I think I understand what the customer wants, so I will respond. Or it's like, ‘hmmm, I think we are at the point where I better escalate, this seems like a big deal and maybe I want to pass that over to an account executive’. Right? So that's really where it becomes really interesting, right? It's like the ability to autonomously decide reason ‘yes, I'll continue to engage by email or I will escalate it’. And so to me, you know, that's where really the potential is.
EM: Yeah, one of the things that we talked about a little bit last month was how quickly people came around to the idea that there was really a ‘there there’ with Agentic AI, you know, you had compared it to when the real sort of next gen GenAI systems were coming out. People very quickly groked that like, this really is a game changing thing. Now that we're sort of like a month on from there, you've had more conversations with folks who are using the developer edition, do you feel like people are just as excited, just as enthused? Do you feel like people are inclined to be skeptical still because of the kind of buzz wordiness of the term? How is that kind of feeling?
CC: Well, I think initially, if you look at the industry, it's like we always need a new shiny object, right? When we are in this world, and there have been things that were very shiny and didn't last long. The way I've been thinking about my own career, I don't know if I used that analogy last time we spoke, but I've always thought about it as: I'm a surfer in the ocean. And what's tricky about being a developer is that you always need to catch the right next wave. That's just like a surfer. But there are many waves and there are waves that are going to take you nowhere, and there are waves that are going to take you far. And I've always really, for many years, I've thought about my career as a developer that way. Because there is always that new shiny framework or that new shiny… I mean, the opportunities to learn new things in front of you are limitless -
EM: Mm-hmm [agreement]
CC: And it's easy to get sucked into that process of going on the wrong path. And then you didn't catch the right wave. As a developer, you always need to be a little skeptical about the new framework, the new, you know, the new language, a new platform, because that's a healthy kind of thing to do because otherwise it's easy to get on the wrong path.
BP: You're sitting out there and you see the crypto wave coming. Is this the right-? The VR waves coming -
EM: Do I paddle for this one? [laughter] –
BP: Yeah. You know, it's like very different end results depending on which wave you choose. Very different -
CC: Exactly. And, and your time is limited, right. So you can't really kind of learn them all. Or at the very beginning when we started talking about Generative AI and then Agentic AI, people were like ‘eh, okay, yeah, I'm not sure. We'll see’. You know, there have been other things like that. But you know, like you said, Eira very, very quickly people realized ‘Wow’, because -
EM: Mm-hmm [agreement]
CC: You know, when you see it, when you experience it. Seeing it is the first step but it's easy to, you know, we all know it's easy to build a demo that looks really amazing. But when people experience it, it's really hard not to see the power and really how it changes what's possible. And like I said, that idea of writing the building blocks that you kind of feed to the LLM and see what the LLM can do with it. And then, from there, like I said, the possibilities are endless. LLMs can solve problems that you didn't even anticipate. And then you have a ton of questions as a developer. It's like, ‘okay, well I knew how to build UIs, for example, for very defined problems. How do I build a UI when I don't even know the problems that my solution can solve?’ Because a UI is typically limited to a number of things that you anticipated. And so, the answer to that is, well, it's conversational. But is the future of software really just going to be text or voice? Like, you're gonna ask it. There are times when a rich UI, you know, makes a ton of sense, right? Because it's a better way to visualize data or to interact with the data. So, a lot of time today you are seeing, right now you're still seeing a lot of examples where the main application is a rich UI. And then you have that little chat box looking thing where you can ask your questions. I think what's starting to happen is the exact opposite, right? Where the conversation becomes the main UI and then you see text, but like I said, text is not always the best way to interact with your data. So now you start seeing rich UIs in the flow of the conversation, in support of the conversation as opposed to the other way around, which was, ‘okay, I have my rich UI app, and then here on the side, bottom right, you can ask a few questions’. So you had the conversation in support of the rich UI. Now it's really the rich UI in support of the conversation. You know, these are all really interesting things that I think developers are starting to see. That's why, for example, using Slack as the front end for the conversation because that's a conversational tool that people are using every day. It's, for example, a great option and then in the flow of that conversation, having rich UI components, I think is really fascinating in terms of the potential.
BP: Yeah, it is really interesting because the AIs are getting increasingly multimodal and they might decide, to your point, that there's a way to answer the question that makes more sense as a rich 3D point cloud as opposed to a text box.
CC: Exactly. And you know, the important thing to make that happen is that you need data about your data, right? So, if the response to a question is data, but now you want to represent the data in a more interesting way, in a more actionable way, you need information about that data to know how to represent it. And that's essentially metadata, right? And that's why we put a lot of emphasis on metadata on the platform, which is really that information about the data. The data about the data, so that if we decide to represent it in some kind of a visual component, in a chart, in a web component, we have enough information to actually build that rich component on the fly using metadata. And I think that's going to be super important as well because we often speak about the importance of data in AI. I think the importance of metadata is also very important.
EM: It's that time of the show where we normally shout out somebody who has won a lifeboat badge for coming on Stack Overflow exchange and sharing some wisdom. But today I wanted to put the question back out to the community. If you are a developer who is working with AI agents, what kind of stuff are you building? And what kind of tools are you looking for? What kind of problems are you interested in solving with AI tools? Send us an email at podcast@stackoverflow.com and let us know what you're working on. I have been Eira May, I am the Business Editor at Stack Overflow, and you can always find me on LinkedIn.
BP: Hey everybody. I'm Ben Popper. I'm one of the hosts of the podcast here at Stack Overflow. Find me on X at Ben Popper, or yeah, hit me up on LinkedIn.
CC: Yes. So I'm Christophe Coenraets. I run Developer Relations at Salesforce. You can find me on LinkedIn at Christophe Coenraets. And, you know, we spoke a lot about the developer edition, so you can all get it. And the place to get it is developer.salesforce.com/signup. And that's the place for you to sign up for that developer edition.
EM: Yeah, and let us know what you start building once you've signed up! Thanks everybody. We'll talk to you next time.