
Sign up to save your podcasts
Or
Today on the Salesforce Admins Podcast, we talk to Joshua Birk, Admin Evangelist at Salesforce.
Join us as we chat about why your metadata is crucial for building effective AI agents. You should subscribe for the full episode, but here are a few takeaways from our conversation with Joshua Birk.
As the Admin Evangelist team has been helping people get started with Agentforce, we’ve noticed that the key to unlocking this new technology is to revisit some of the oldest concepts about the Salesforce platform.
That’s why I brought Josh Birk on the pod to talk about metadata and multi-tenant architecture. If you need a refresher, that’s the idea that Salesforce is like an apartment building where each org is an apartment. Your stuff is in your individual unit, but the entire building shares resources like water and electricity.
So what’s the difference from 2010? As Josh explains, it’s that every apartment comes standard with an Agentforce-powered robot butler.
Imagine you’re sitting down for dinner, and you want your robot butler to set the table—how does it know where the forks are? And what happens if they’re buried in your junk drawer?
Clearly, a robot butler will be more helpful if you keep your apartment organized. And, as Josh points out, the same is true for your Salesforce org. AI agents rely on your metadata, like description fields and field types, to help them respond correctly and find what your users are looking for.
With longstanding orgs, there can be an “if it ain’t broke, don’t fix it” mindset, but that’s the equivalent of throwing everything in the junk drawer. Doing a little spring cleaning and organizing your metadata helps Agentforce help you.
The key thing Josh wants you to realize is that you’re already an AI builder. An agent is just another user in your org, and so the work you do to make your data easy to use is also what powers the solutions you build in Agentforce.
That’s why it’s so important to fall back on Salesforce fundamentals. Building an agent is the easy part. The hard part is making sure your metadata is in a good place to support your AI solutions, but that’s the work that admins do every day.
There’s so much more great stuff from Josh in this episode, so be sure to take a listen. And don’t forget to subscribe to the Salesforce Admins Podcast to catch us every Thursday.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
First of all, I think it’s one of those things that does get bandied about, but sometimes it doesn’t stick in people’s heads. For instance, I was talking to somebody about Trailhead, and they were like, “Well, is that why we have multi-tenancy?” And I’m like, “No, no, no, no, no. Multi-tenancy was the thing, that was day one. That was the magic fuel that was moving our product before anything else was getting built on top of it.”
And I think it’s important to kind of have a basic understanding, because every now and then, as a developer, one of the things that people, if you come from Java, or you come from C++, and one of the things you’re used to doing is whatever you want, if you want to spin up a crazy thread in an application and have it run for three days straight, the only person stopping you from doing that is you. And so then people come into Salesforce platform and they’re like, “Well, Apex has these limits to it. You can’t spin up too many CPU cycles, you can’t do too many searches.” You can’t do these things. And some developers are like, “Why are you tying my hands back behind my back?” And it’s like, because part of the multi-tenant architecture makes sure that Mike’s org might be sitting next to my org in terms of hardware. So we’re sharing a CPU.
And if Mike misbehaves too much, you might steal CPU cycles from me, and vice versa. You might slow down my application by no intent of yours. You don’t even know I’m on the other side of the wall, but you’re using up all of my power. And our multi-tenant architecture basically makes sure that doesn’t happen. And it makes sure that you can go in and do all the things that we give you the rules to, and I can go and do all the things we get the rules to, and we’re always going to be play safe with each other.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
And then start breaking all it down to the stuff that you were just saying. Your pictures are your own, the boxes you have in the attic, they’re all your own. All of that is, so this is why trust is our number one value. You trust us to hold your apartment near and dear to you, and you trust us to keep that security layer, you trust us to keep the powers on. All of that is the same stuff that the agents are building on top of. That is why they’re a very easy to use enterprise solution for AI, because they’re like the robot butler who lives in your apartment. And all of that other stuff, the security, the power, the storage, all of that stuff, it just comes with the apartment. And your agent is able to understand where all your pictures are, and where all your boxes are, and how those boxes are structured.
And one analogy I like to give is like, well, if you’re going to ask your robot butler, aka agent force, for a fork, if I’m a guest in your house, like, “Hey, Mike, can I get a fork?” You would know which drawer to tell me. You would be like, “Go to the second drawer next to the stove. That’s where the forks are.” How does the agent know that? And the agent knows that because of metadata. And metadata is the second tier of what makes multi-tenancy work.
One of the things back in my workshop days, a question got a lot, especially because I’m talking to old school Oracle developers and Java developers, and they’re like, “Well, what’s your web stack?” What they’re asking is, are you running Java? Are you running Oracle? And stuff like that. I’m like, “I could answer this for you, but it’s not going to be the answer you think it is, because our data structure doesn’t start like a normal database does. Our database structure starts with multi-tenancy with a metadata tier on top of it.” And what that metadata allows you to do is you can put your forks in a drawer, I can put my forks in a different drawer, and metadata can tell the agent, “This is where those forks are held.”
And let’s say you ripped your kitchen apart. Now, if you’re going to rip your kitchen apart and put your kitchen back together again, you’re going to get a contractor. You’re not just going to go pay a couple of teenagers with some sledgehammers and just knock everything down. I mean, it might start that way, it could be entertaining.
Mike:
Josh Birk:
Mike:
Josh Birk:
So if there’s any ambiguity, because sometimes you might have a similarly named field, you might have a similarly named object, something like that, and you want to make sure people are guided to the right drawer, so find that fork, well, if it’s a good idea for a human, it’s a great idea for an agent, because when you say, “Get me the largest fork in the drawer,” it’s going to go through what it knows about the data model. And if it sees the word fork in a description, it knows it’s going in the right direction. You’re giving it that breadcrumbs.
And this is why it’s like, “Which descriptions fields do you mean, Josh?” And it’s all of them. It’s the ones in the custom actions, and we won’t even get into instructions yet. It’s all the ones in your custom fields, it’s all the ones in your custom objects, et cetera, et cetera, et cetera. You want to leave that little breadcrumb trail so that when agent force is doing its little agentic reasoning, and it’s like, “Which custom action should I use? Which standard action should I use?” You’re going to end up with a much higher rate of success.
Mike:
Josh Birk:
Mike:
Josh Birk:
Now, people might be listening to this like, “Well, okay, I understand that, but how powerful can a description field be?” And it’s true. It’s only so powerful, but the great thing is, this is why, and it’s one of the things I think people, I really like talking about this, because one of the things is I think it demystifies agents a certain extent. We treat agents almost like, I just referred to it as a robot butler. So we’re already into a sci-fi. We’ve gone from an apartment complex to Star Wars, right?
Mike:
Josh Birk:
But if you say the word my without giving any context to the agent, it’s going to think you mean the current owner. But instructions, you can say, “Hey, in this topic, when I say my, when I say blog, I mean these different things.” And so that’s the first place that the breadcrumb trail is going to start going, and it can be one of the little behavioral issues. It’s not responding in the right way, or it’s not using the right style and things like that. The prompt builder instructions can solve so much just by writing a few sentences.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Now, going back to the blueprint and the description fields and the topics, and also the other thing to mention here is the quality of your data. How well-structured is your kitchen when it comes to, do you have all the forks in one drawer, and the spoons in another drawer? Do you have one of those nice slotted trays?
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
So, I’ll go way off on a tangent, but I swear I’m getting somewhere. Back when I worked for a major retailer as their lead client side developer, we had a problem with pop-up windows. And the problem was, because everybody knows-
Mike:
Josh Birk:
Mike:
Josh Birk:
They had built their own feature without even asking us. And so we fixed it by adding that functionality in, like the way we wanted it to work, and everybody was really, really happy. But that’s the kind of thing that happens within a Salesforce org. You didn’t think that you gave them the feature of transcribing their phone calls into the long text field that was meant for something else, but you did, and they’re using it. And that’s the kind of, I think you’re credited for saying the walk around the broom style of admin.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
And I was like, “Oh, it just dawned on me,” because I had just done a Salesforce deployment. I was like, “I didn’t look at that page that I deployed. I looked at it from my desk with my headphones on, uninterrupted, knowing what I was looking for, not on a phone call with an irate customer because their plant had just died or something, and trying to work through how do I disseminate this information.” So there you go. There’s the history of Sabla.
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
So, if you don’t even have to be Jen League, Queen of Flow, you don’t have to be a crazy flow-natic. You just need to know how to build flow. Now, you know how to build flow, great: you know how to build custom actions. And when you take that flow and you convert it over to a custom action, all of that work that you’ve put into the apartment leading up to that moment is going to make that agent so much more successful. So it’s like, don’t think of it as this brand new thing that’s off on its own little lonely island, it’s more similar to when we released Flow, or when we released things that are new, but they’re new on top of the platform. And so all of the things you already know how to do on the platform, all of those things are going to be important.
In one of the blog posts we have coming out, one of the things Jen reminded me to put in there is, “Remember to think of your agent like a user. It’s a Rosey. It’s another thing in your apartment running around doing stuff.” And so all of your ability to be like, “Oh, what fields should I have access to? What permission sets should I give it?” All of that, it’s all stuff that’s going to make your agent more successful.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
The post Use Metadata To Empower Salesforce Agents appeared first on Salesforce Admins.
4.7
200200 ratings
Today on the Salesforce Admins Podcast, we talk to Joshua Birk, Admin Evangelist at Salesforce.
Join us as we chat about why your metadata is crucial for building effective AI agents. You should subscribe for the full episode, but here are a few takeaways from our conversation with Joshua Birk.
As the Admin Evangelist team has been helping people get started with Agentforce, we’ve noticed that the key to unlocking this new technology is to revisit some of the oldest concepts about the Salesforce platform.
That’s why I brought Josh Birk on the pod to talk about metadata and multi-tenant architecture. If you need a refresher, that’s the idea that Salesforce is like an apartment building where each org is an apartment. Your stuff is in your individual unit, but the entire building shares resources like water and electricity.
So what’s the difference from 2010? As Josh explains, it’s that every apartment comes standard with an Agentforce-powered robot butler.
Imagine you’re sitting down for dinner, and you want your robot butler to set the table—how does it know where the forks are? And what happens if they’re buried in your junk drawer?
Clearly, a robot butler will be more helpful if you keep your apartment organized. And, as Josh points out, the same is true for your Salesforce org. AI agents rely on your metadata, like description fields and field types, to help them respond correctly and find what your users are looking for.
With longstanding orgs, there can be an “if it ain’t broke, don’t fix it” mindset, but that’s the equivalent of throwing everything in the junk drawer. Doing a little spring cleaning and organizing your metadata helps Agentforce help you.
The key thing Josh wants you to realize is that you’re already an AI builder. An agent is just another user in your org, and so the work you do to make your data easy to use is also what powers the solutions you build in Agentforce.
That’s why it’s so important to fall back on Salesforce fundamentals. Building an agent is the easy part. The hard part is making sure your metadata is in a good place to support your AI solutions, but that’s the work that admins do every day.
There’s so much more great stuff from Josh in this episode, so be sure to take a listen. And don’t forget to subscribe to the Salesforce Admins Podcast to catch us every Thursday.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
First of all, I think it’s one of those things that does get bandied about, but sometimes it doesn’t stick in people’s heads. For instance, I was talking to somebody about Trailhead, and they were like, “Well, is that why we have multi-tenancy?” And I’m like, “No, no, no, no, no. Multi-tenancy was the thing, that was day one. That was the magic fuel that was moving our product before anything else was getting built on top of it.”
And I think it’s important to kind of have a basic understanding, because every now and then, as a developer, one of the things that people, if you come from Java, or you come from C++, and one of the things you’re used to doing is whatever you want, if you want to spin up a crazy thread in an application and have it run for three days straight, the only person stopping you from doing that is you. And so then people come into Salesforce platform and they’re like, “Well, Apex has these limits to it. You can’t spin up too many CPU cycles, you can’t do too many searches.” You can’t do these things. And some developers are like, “Why are you tying my hands back behind my back?” And it’s like, because part of the multi-tenant architecture makes sure that Mike’s org might be sitting next to my org in terms of hardware. So we’re sharing a CPU.
And if Mike misbehaves too much, you might steal CPU cycles from me, and vice versa. You might slow down my application by no intent of yours. You don’t even know I’m on the other side of the wall, but you’re using up all of my power. And our multi-tenant architecture basically makes sure that doesn’t happen. And it makes sure that you can go in and do all the things that we give you the rules to, and I can go and do all the things we get the rules to, and we’re always going to be play safe with each other.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
And then start breaking all it down to the stuff that you were just saying. Your pictures are your own, the boxes you have in the attic, they’re all your own. All of that is, so this is why trust is our number one value. You trust us to hold your apartment near and dear to you, and you trust us to keep that security layer, you trust us to keep the powers on. All of that is the same stuff that the agents are building on top of. That is why they’re a very easy to use enterprise solution for AI, because they’re like the robot butler who lives in your apartment. And all of that other stuff, the security, the power, the storage, all of that stuff, it just comes with the apartment. And your agent is able to understand where all your pictures are, and where all your boxes are, and how those boxes are structured.
And one analogy I like to give is like, well, if you’re going to ask your robot butler, aka agent force, for a fork, if I’m a guest in your house, like, “Hey, Mike, can I get a fork?” You would know which drawer to tell me. You would be like, “Go to the second drawer next to the stove. That’s where the forks are.” How does the agent know that? And the agent knows that because of metadata. And metadata is the second tier of what makes multi-tenancy work.
One of the things back in my workshop days, a question got a lot, especially because I’m talking to old school Oracle developers and Java developers, and they’re like, “Well, what’s your web stack?” What they’re asking is, are you running Java? Are you running Oracle? And stuff like that. I’m like, “I could answer this for you, but it’s not going to be the answer you think it is, because our data structure doesn’t start like a normal database does. Our database structure starts with multi-tenancy with a metadata tier on top of it.” And what that metadata allows you to do is you can put your forks in a drawer, I can put my forks in a different drawer, and metadata can tell the agent, “This is where those forks are held.”
And let’s say you ripped your kitchen apart. Now, if you’re going to rip your kitchen apart and put your kitchen back together again, you’re going to get a contractor. You’re not just going to go pay a couple of teenagers with some sledgehammers and just knock everything down. I mean, it might start that way, it could be entertaining.
Mike:
Josh Birk:
Mike:
Josh Birk:
So if there’s any ambiguity, because sometimes you might have a similarly named field, you might have a similarly named object, something like that, and you want to make sure people are guided to the right drawer, so find that fork, well, if it’s a good idea for a human, it’s a great idea for an agent, because when you say, “Get me the largest fork in the drawer,” it’s going to go through what it knows about the data model. And if it sees the word fork in a description, it knows it’s going in the right direction. You’re giving it that breadcrumbs.
And this is why it’s like, “Which descriptions fields do you mean, Josh?” And it’s all of them. It’s the ones in the custom actions, and we won’t even get into instructions yet. It’s all the ones in your custom fields, it’s all the ones in your custom objects, et cetera, et cetera, et cetera. You want to leave that little breadcrumb trail so that when agent force is doing its little agentic reasoning, and it’s like, “Which custom action should I use? Which standard action should I use?” You’re going to end up with a much higher rate of success.
Mike:
Josh Birk:
Mike:
Josh Birk:
Now, people might be listening to this like, “Well, okay, I understand that, but how powerful can a description field be?” And it’s true. It’s only so powerful, but the great thing is, this is why, and it’s one of the things I think people, I really like talking about this, because one of the things is I think it demystifies agents a certain extent. We treat agents almost like, I just referred to it as a robot butler. So we’re already into a sci-fi. We’ve gone from an apartment complex to Star Wars, right?
Mike:
Josh Birk:
But if you say the word my without giving any context to the agent, it’s going to think you mean the current owner. But instructions, you can say, “Hey, in this topic, when I say my, when I say blog, I mean these different things.” And so that’s the first place that the breadcrumb trail is going to start going, and it can be one of the little behavioral issues. It’s not responding in the right way, or it’s not using the right style and things like that. The prompt builder instructions can solve so much just by writing a few sentences.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Now, going back to the blueprint and the description fields and the topics, and also the other thing to mention here is the quality of your data. How well-structured is your kitchen when it comes to, do you have all the forks in one drawer, and the spoons in another drawer? Do you have one of those nice slotted trays?
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
So, I’ll go way off on a tangent, but I swear I’m getting somewhere. Back when I worked for a major retailer as their lead client side developer, we had a problem with pop-up windows. And the problem was, because everybody knows-
Mike:
Josh Birk:
Mike:
Josh Birk:
They had built their own feature without even asking us. And so we fixed it by adding that functionality in, like the way we wanted it to work, and everybody was really, really happy. But that’s the kind of thing that happens within a Salesforce org. You didn’t think that you gave them the feature of transcribing their phone calls into the long text field that was meant for something else, but you did, and they’re using it. And that’s the kind of, I think you’re credited for saying the walk around the broom style of admin.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
And I was like, “Oh, it just dawned on me,” because I had just done a Salesforce deployment. I was like, “I didn’t look at that page that I deployed. I looked at it from my desk with my headphones on, uninterrupted, knowing what I was looking for, not on a phone call with an irate customer because their plant had just died or something, and trying to work through how do I disseminate this information.” So there you go. There’s the history of Sabla.
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
So, if you don’t even have to be Jen League, Queen of Flow, you don’t have to be a crazy flow-natic. You just need to know how to build flow. Now, you know how to build flow, great: you know how to build custom actions. And when you take that flow and you convert it over to a custom action, all of that work that you’ve put into the apartment leading up to that moment is going to make that agent so much more successful. So it’s like, don’t think of it as this brand new thing that’s off on its own little lonely island, it’s more similar to when we released Flow, or when we released things that are new, but they’re new on top of the platform. And so all of the things you already know how to do on the platform, all of those things are going to be important.
In one of the blog posts we have coming out, one of the things Jen reminded me to put in there is, “Remember to think of your agent like a user. It’s a Rosey. It’s another thing in your apartment running around doing stuff.” And so all of your ability to be like, “Oh, what fields should I have access to? What permission sets should I give it?” All of that, it’s all stuff that’s going to make your agent more successful.
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
Josh Birk:
Mike:
The post Use Metadata To Empower Salesforce Agents appeared first on Salesforce Admins.
4,320 Listeners
1,192 Listeners
3,173 Listeners
427 Listeners
3,979 Listeners
3,684 Listeners
1,012 Listeners
5,920 Listeners
2,941 Listeners
5,255 Listeners
228 Listeners
22 Listeners
234 Listeners
1 Listeners
963 Listeners