Agentforce NOW codeLive: Build Agentic Data 360 Experiences Headlessly — Transcript
Full transcript
- 0:00Hello everyone. Thank you so much for
- 0:02joining Agent Force now Code Live. My
- 0:06name is Mohit. I'm a developer advocate
- 0:08here at Salesforce and I welcome you all
- 0:11to this Code Live episode. Today we're
- 0:13going to be
- 0:15building agentic data experiences, more
- 0:18importantly headlessly, all through
- 0:20coding agents.
- 0:22Uh and I have with me uh the product
- 0:25manager who is sort of working on all of
- 0:28this. Anush, welcome to Code Live.
- 0:31>> Yeah, thank you Mohit. Thank you for
- 0:32inviting me.
- 0:33Uh hi everyone. My name is Anush. I'm
- 0:36the product manager for headless data
- 0:37360 starting with the data 360 MCP
- 0:41server.
- 0:42So I'm really excited to put this
- 0:43hands-on session with Mohit to get you
- 0:45guys started and to see your experience
- 0:47using the MCP server.
- 0:49>> Perfect. Perfect. And and I know Anush
- 0:51you're very busy building these MCP
- 0:53servers, so I really appreciate you
- 0:56taking time today. And for all those of
- 0:58you who are tuning
- 1:00um remember that we also have chats
- 1:02there. So as we sort of show these
- 1:04capabilities to you, if you have any
- 1:07questions, please make use of those chat
- 1:11and um shoot us to us and we'll we'll
- 1:13get back to those questions as they sort
- 1:15of come along. Uh with that uh you know,
- 1:18because this is Salesforce presentation,
- 1:21um
- 1:22I'll put the presentations here, but
- 1:24more importantly uh our forward-looking
- 1:26statement. Uh so please do not make your
- 1:29purchasing decisions based on features
- 1:33um that are uh not available in the
- 1:35product yet or not GA.
- 1:37With that out of our way, again um you
- 1:40know, myself and Anush we're going to be
- 1:41talking about building agentic data 360
- 1:45experiences headlessly.
- 1:47And I just want to remind everyone that
- 1:50uh you know, there is a QR code. I've
- 1:52also uh sort of placed this QR code on
- 1:54your screens on your right. So um
- 1:57make sure that if you have any feedback
- 1:59about this code live or any code live in
- 2:02general, just use that QR code and tell
- 2:06us. With that,
- 2:08how about Anush, we actually take some
- 2:11time and explain
- 2:13to everyone
- 2:15like what is this data 360 headless all
- 2:17about, what is this momentum about, what
- 2:19are these MCP servers. So, let's take
- 2:21some time everyone to get to the same
- 2:23page here.
- 2:24>> Yes.
- 2:25I think that'd be a good start.
- 2:27Starting with what are we doing here?
- 2:29If you started earlier this year using
- 2:31data 360, maybe even in Jan,
- 2:34a lot of the functionality was in the
- 2:37lightning UI in Salesforce.
- 2:39So, even if you're using agents back
- 2:41then, you had to move between the UI,
- 2:44maybe ask a coding agent what you need
- 2:46to do, look at the documentation. So,
- 2:48this led to like some sort of context
- 2:50mismatch between where you're doing work
- 2:52and where you need to go gather this
- 2:53context for.
- 2:55So, in order to solve that, we are
- 2:58starting a really focused headless
- 3:00program.
- 3:02What does this mean?
- 3:04The foundation of this program for us is
- 3:06making sure all our functionality in the
- 3:09UI is available through APIs. So, that
- 3:11is the foundation layer in which we are
- 3:14building.
- 3:15Once those APIs are exposed, we have now
- 3:18released the data 360 MCP server.
- 3:22Actually, earlier in June we had a
- 3:23developer preview of a local MCP server
- 3:26>> Right.
- 3:26>> we got a lot of feedback about how our
- 3:28users are using it. We saw the user
- 3:31experience with a lot of great feedback
- 3:33from that.
- 3:34And from that we realized that the
- 3:36hosted remote MCP server was a better
- 3:38option for most users.
- 3:41>> Perfect.
- 3:42>> Now,
- 3:44why is this really important, right?
- 3:47It's not just about you going to Claude
- 3:49or going to Codex and telling uh
- 3:51Codex, go do this thing or Claude do
- 3:53this thing no make no mistakes. It's not
- 3:55just that. It's about providing trusted
- 3:57context
- 3:59and also with governed resources and
- 4:03pre-built skills via the MCP server.
- 4:06Why is this important?
- 4:09So, if you notice
- 4:10agents
- 4:12are good at solving things but they may
- 4:15also not be the best at following the
- 4:17shortest path to a solution.
- 4:20But you being Salesforce developers or
- 4:22industry experts know what is the best
- 4:24path.
- 4:25So, this MCP server paired with
- 4:28specialized skills reduce the number of
- 4:31tokens required for an agent to go from
- 4:33zero to one.
- 4:34So if you notice the MCP server today,
- 4:37we have just three tools.
- 4:39We have the search tool.
- 4:42We have payload examples and we have
- 4:44execute.
- 4:46Why just three tools, right?
- 4:49We notice that with dynamic tool
- 4:51loading, when you ask for a certain
- 4:54intent the agent can then then search
- 4:57for what what it needs. For example, it
- 4:59could be segmentation. You want to
- 5:01create a user segment to maybe
- 5:03uh
- 5:04target a specific user segment. And then
- 5:07the agent finds the exact payload
- 5:08examples for that tool. And that reduces
- 5:11the number of tones an agent needs
- 5:14to successfully solve this.
- 5:16So, which is why we have followed this
- 5:17dynamic tool loading method as well as
- 5:20coming soon will be skills accompanied
- 5:23with the MCP server.
- 5:24This will reduce the number of tones an
- 5:26agent needs, the number of tokens you
- 5:28need, and also it provides best
- 5:30practices that practitioners using the
- 5:33Data 360 platform used to build it.
- 5:36S-
- 5:38So now, not just can you do this through
- 5:40Claude or Codex, but you get this really
- 5:42smart reasoning engine paired with you
- 5:45following the best practices and
- 5:47executing on it. Which is why I'm so
- 5:48excited for this.
- 5:50>> Awesome. Awesome. No, this is this is
- 5:52great. Um one thing I really want to
- 5:54call out is look at the simplicity of
- 5:57this whole design, right? Uh when I when
- 6:00I was sort of um
- 6:02thinking about, you know, these MCPs for
- 6:05Data360 and myself, you know, um
- 6:08I'm I I can call myself Data360 uh
- 6:12developer or
- 6:13>> you can.
- 6:14>> you know, Data Blazer. Um
- 6:16so, there are there are so many tools in
- 6:19Data360, as as you know, um Anush, like
- 6:23there's so much of power that this
- 6:25Data360 has and the tools it has. You
- 6:28know, let's talk about, you know,
- 6:29creating these data streams to bring
- 6:31data from like, you know, multiple
- 6:33system, like thousands of systems,
- 6:35right? You can ingest data through
- 6:37various mechanisms, like, for example,
- 6:39zero copy is one of my favorite. Because
- 6:42if you have a Snowflake
- 6:43>> Yeah.
- 6:45>> you know, sort of zero copy uh data into
- 6:47Data360.
- 6:49Uh but still like all of those involve a
- 6:51lot of clicks. And when I thought about
- 6:53MCPs for, you know, Data360
- 6:56with all these tools, I was like, this
- 6:58is going to be like, you know, maybe n
- 7:00number of tools, right? Thousands of
- 7:02tools here in this
- 7:03>> Yeah.
- 7:03>> in this scenario because there are
- 7:05thousands of APIs. But uh you know, when
- 7:07I saw like, okay, you know, you guys are
- 7:09like only exposing like only three
- 7:12different MCP tools,
- 7:14um so, that's again less number of token
- 7:17consumptions and also much easier for
- 7:20the AI to like sort of reason, right?
- 7:22Because it's now progressively sort of
- 7:24progressing towards what what is
- 7:26required than, you know, just loading
- 7:29everything in one shot and sort of
- 7:31consuming the context window and then
- 7:33sort of having AI just hallucinate it,
- 7:35right?
- 7:35>> Yeah.
- 7:37>> Perfect. Perfect. Okay. Uh so, for all
- 7:39of you, right? Um today what I what what
- 7:41we are thinking is we're going to make
- 7:43this much more like practical just
- 7:46showing the use of it. Um, and uh if you
- 7:49are new to Data 360, welcome. Um, no
- 7:52issues. We're we'll try to like cover
- 7:54those aspects as well. Like what is Data
- 7:56360? What does it do? Etc.
- 7:59Um, but I have a business use case and I
- 8:02think, you know, all of us are there to
- 8:04solve business problems, right? So, I
- 8:06thought it will resonate well if we
- 8:08start with the the use case here.
- 8:10And uh Anush, for us today the use case
- 8:13is we have this Pronto, a food delivery
- 8:15platform. And this is a sample
- 8:17application that developer advocacy team
- 8:19sort of uses it uh to showcase uh
- 8:22different capabilities of the platform.
- 8:25So, we are taking that Pronto uh
- 8:27platform today and we will leverage the
- 8:29Data 360 as a centralized system for
- 8:32context for agents that Pronto is
- 8:35building. Um, basically it's going to be
- 8:38unifying data from multiple sources,
- 8:41activating that data across different
- 8:43applications and different agents. So,
- 8:45this is the primary use case of Data 360
- 8:48in case you are wondering what is Data
- 8:49360, right? Is that context engine that
- 8:52will drive agents, uh your apps, you
- 8:55know, think of it like a big data
- 8:57processing engine, right? With all the
- 8:58tools and batteries included so you can
- 9:01quickly activate, you know, the data um
- 9:04for context as well as for um sort of
- 9:08building applications if if that's what
- 9:10you're doing with this data, right? Um,
- 9:12so that's so our use case, our excite
- 9:14business problem that we're going to
- 9:16solve we're going to try to solve in
- 9:17this session. Although, I don't know how
- 9:19far we will get along with these coding
- 9:21agents and sort of working because, you
- 9:23know, building something live is like,
- 9:26you know, hours of work. Like even
- 9:27clicking through it. And agents is going
- 9:29to simplify a lot of it. We'll try to to
- 9:31sort of do a lot of stuff here um live.
- 9:34So, really what we are trying to solve
- 9:36is uh for Pronto, the customer is
- 9:39showing up now in as two different
- 9:41peoples. Even though it's same customer,
- 9:44um but it's showing up as two different
- 9:46people because uh you know,
- 9:49one way the information is coming
- 9:51through their delivery app system, which
- 9:54uses uh basically an application which
- 9:57uses the the email to order, and then
- 10:00they are collecting all the reviews in
- 10:02Salesforce CRM.
- 10:04And we need to able to sort of activate
- 10:07it using data from multiple of these
- 10:10systems. The rules that we're going to
- 10:12write is not just belongs to the reviews
- 10:15here. Like our rules here, as you can
- 10:17see, is two different rules here. One is
- 10:19if the order da- date is sort of lapsed
- 10:22um 60 days beyond, and also the rating
- 10:26or the rating is less than two. So you
- 10:27you can see we're creating a
- 10:29segmentation operation to sort of win
- 10:32back these customers from multiple
- 10:34systems. And these are extra type of
- 10:36problems actually uh at scale that
- 10:38Data3Sixty sort of solves.
- 10:40>> Yeah, so this is a common problem I've
- 10:42seen Mohit where a lot of customers may
- 10:45have unhappy customers they're serving.
- 10:48And how do you how do you figure out
- 10:49that's the right customer, right? When
- 10:50you have this data among multiple
- 10:52different systems, how do you know
- 10:54this reviewer is actual customer of you?
- 10:56Is that a real review? Is it a fake
- 10:57review? How do you figure out should you
- 10:59reach out to them or is it worth your
- 11:01time?
- 11:02>> Perfect, perfect. Yeah, yeah. No, this
- 11:03is the use case. I mean used to call
- 11:05this CDP use case. I I remember the term
- 11:08CDP use case, right?
- 11:11But here's here's the simple design to
- 11:13solve these kind of things, and I just
- 11:15put it out here because this is very
- 11:17common um you know, solution design that
- 11:21you um that you as solution engineers
- 11:24are are doing most of the time. So you
- 11:26will basically in Data3Sixty create data
- 11:29streams
- 11:31uh for different data um
- 11:34sources. In this case, it is that app,
- 11:37the data coming from the app,
- 11:39which in my case I've simplified it
- 11:41actually with just a CSV file. But,
- 11:43think about it's coming from your mobile
- 11:45app or something. And then
- 11:48the data the other data
- 11:51source here is the Salesforce itself
- 11:52that has all the reviews. So, we take
- 11:54all those so we create data streams, we
- 11:56create
- 11:58what we call data lake objects,
- 12:01and then from that we create data model
- 12:05objects by mapping them to the, you
- 12:07know, the canonical data model that
- 12:10Data 360 provides. So, Data 360 by
- 12:12default has pre-built data models for
- 12:16different business purposes
- 12:18to activate that data
- 12:20or unify that data. Right? So, so that
- 12:23is unification. So, we'll do that
- 12:24unification. Generally, that that is
- 12:26your workflow, creating streams and
- 12:29doing this unification or mapping it to
- 12:31the right canonical data model within
- 12:34the Data 360. And then you resolve using
- 12:36identity resolution. In this case, it's
- 12:38very simple like I've just taken email
- 12:40as that common thing. But, it could be,
- 12:43you know, multiple logic like multiple
- 12:44stuff. It could be very complex too.
- 12:47Fuzzy matches, etc. So, based on that
- 12:49you map them and match it and create
- 12:52that one single identity record
- 12:55from which we can create this
- 12:57segmentation which will be another
- 13:00data object called win-back diners,
- 13:02you know, where we sort of segment all
- 13:05those customers who have like sort of
- 13:07not ordered from Pronto app.
- 13:11You know, it's been like 60 days since
- 13:13they ordered or
- 13:15the rating has like sort of fallen less
- 13:17than two.
- 13:19So, this is the solution design. Now
- 13:22you know, because I've worked with the
- 13:24system, I know the solution design.
- 13:27but when you are solving a problem,
- 13:29right? And because we have different
- 13:31tool sets,
- 13:32another way to approach this kind of
- 13:35problem these days with the agents is
- 13:39ask the agent itself, like how do I
- 13:42solve using that MCP that we have, and
- 13:45the agent will definitely sort of look
- 13:47at all the payloads and can recommend
- 13:49that these are the things. Again, in
- 13:51this loop you still need to be in the
- 13:53loop because
- 13:54you need to judge whether whatever that
- 13:56agent is sort of telling you is correct
- 13:58or not. So, you still need to understand
- 14:00all of this. Uh
- 14:02but uh once you sort of understand, then
- 14:05this whole MCP uh and skills and
- 14:08everything becomes much more useful
- 14:10because then they are like, you know,
- 14:11power tools, which you can put them into
- 14:14action. I can move with a very uh
- 14:17you know, rapid pace. Um any thoughts
- 14:19here, um Anush?
- 14:21>> Yeah, so uh I think you really nailed
- 14:24that one because even for me, when I if
- 14:27for me, sometimes the business problem
- 14:29is just I want to win back these diners.
- 14:31Like, I do not know anything about data
- 14:33360. I don't know I need to unify this
- 14:35data. All I need to do is my goal is I
- 14:37need to win back these customers who
- 14:39have dropped off from the platform,
- 14:40right? And I usually pair with Claude or
- 14:43Codex, and I'm like, "Hey, this is my
- 14:45problem.
- 14:46This is what I have at my disposal.
- 14:48Help me brainstorm through the
- 14:49solution."
- 14:51And as long as you're connected to the
- 14:52tools to the agent, whether it's Claude
- 14:54or or Codex or even GitHub co-pilot,
- 14:58they'll be able to help you reason
- 14:59through your various options. And uh
- 15:02like Mohit mentioned, that since Mohit's
- 15:03a data 360 expert, he has a good grasp
- 15:06of what he needs to do, that is unify,
- 15:09resolve, and segment. But if you are new
- 15:11to data 360, just ask your coding agent,
- 15:14"Hey, this is my problem. How can I
- 15:16solve this?" And you'll be able to
- 15:17brainstorm through it.
- 15:19>> Yeah. Yeah.
- 15:20Yeah, exactly, perfect. Um so, let's
- 15:23let's go into that live coding segment
- 15:25with the coding agent
- 15:27and the data 360 MCP. So, we want to
- 15:30take this MCP server for a ride today
- 15:32and see how it kind of does.
- 15:35And again, you know, we want to be
- 15:37upfront honest with you. So, this is all
- 15:39GA, right?
- 15:40There are certain things in here that we
- 15:43are still bridging like of course data
- 15:45360 has a lot of like a lot of APIs and
- 15:48a lot of capabilities, right? And to
- 15:50bring all of them to the MCP, I think
- 15:53Anush is Anush and the team is working
- 15:55on. But, I would say from my experience
- 15:57working on this now from like last 2
- 15:59weeks, I can say, you know, definitely
- 16:01this is this is
- 16:03helping me like, you know, I'm able to
- 16:05sort of move with the speed of my intent
- 16:08here.
- 16:09You know, just knowing what I need to do
- 16:11and then using this MCP has has been a
- 16:14helper. Anush, do you have any thoughts
- 16:16on what are the things that you're still
- 16:17working on?
- 16:19Yeah.
- 16:19>> So, anything that's a UI functionality,
- 16:24it will be exposed to the API and you
- 16:26will see it in the MCP server.
- 16:28If there are some recent APIs, you
- 16:30should be seeing them shortly in the MCP
- 16:32server.
- 16:33But, there are also certain APIs or
- 16:35certain functionality we might not
- 16:37expose.
- 16:38An example for that is
- 16:40imagine you need to set up a connection
- 16:42which requires putting your credentials
- 16:44into an API or credentials into the UI
- 16:47to connect from system to system.
- 16:49>> Right.
- 16:49>> We might not want you to put the those
- 16:51credentials into the agent because then
- 16:54that might be stored somewhere which
- 16:56might lead to a security lapse.
- 16:59>> Yeah.
- 16:59>> So, you might notice very few sliver of
- 17:03functionality that we might
- 17:05intentionally keep away
- 17:07from the lens of
- 17:10security as well as governance. But,
- 17:12whatever other functionality that you
- 17:14have, you will see it in there. If we If
- 17:17we have missed something, please reach
- 17:18out to us. Let us know if something
- 17:20doesn't work out for you well, let us
- 17:21know and we'll be We'll make sure to
- 17:23address that as soon as possible.
- 17:25>> Yeah, perfect. Uh perfect. Um so first
- 17:28actually, so I want to show you the
- 17:30Pronto instance. So this is the the
- 17:32Pronto instance that I was talking about
- 17:34and uh
- 17:35this is basically what I'll be using for
- 17:37my uh building uh demos. So one of the
- 17:40things we will do is ask the agent to
- 17:42create data streams
- 17:44um and then create identity resolutions
- 17:47today with um data streams from
- 17:49Salesforce
- 17:50um
- 17:51pulling all the reviews. So these are in
- 17:52Salesforce. Now we want to pull it into
- 17:54data data 360 here so we can unify with
- 17:57the app orders data. So that's something
- 17:59we'll be doing.
- 18:00But before all this, there is some setup
- 18:03that is involved in um sort of setting
- 18:07up your org so that that MCP gets
- 18:09activated and you can connect to your
- 18:11coding agent. I want to go through that
- 18:14now uh with you. So uh real quick,
- 18:17right? Um you should find
- 18:20this MCP server in your API catalog. So
- 18:25if you go to API catalog, you will see
- 18:27something like data 360 server. So um
- 18:31again for every org, you will basically
- 18:35go there um and this is going to take
- 18:38some some time. Let me actually refresh.
- 18:40Sometimes it does take some time to
- 18:42load. But really uh what it will have is
- 18:44it will have a toggle. Now there is
- 18:47already a Salesforce CLI command too
- 18:51that the CLI team is exposing. It's in
- 18:53beta at this point through which you can
- 18:56um sort of not have to like go through
- 18:58the UI
- 18:59uh to activate the servers. You can just
- 19:02like
- 19:03do it through the CLI itself. So that
- 19:06way you can script it if you need to.
- 19:08Uh let's see why it's kind of hanging up
- 19:10here.
- 19:12Let's give it a shot once more.
- 19:15This is why I like headless actually.
- 19:20The headless makes it so easy. I don't
- 19:22have to click around and sort of wait
- 19:24for, you know, uh this UI to sort of
- 19:26load up. Uh but while that is loading
- 19:29up, right? Uh once that is sort of
- 19:32activated, which it will, uh it will
- 19:34show up here. But let me show you what
- 19:37you need to do. So so the first step is
- 19:38obviously to activate it. The next step
- 19:41here is to
- 19:43um create something called as external
- 19:46uh
- 19:47client credential app. And uh it's In my
- 19:51case, my coding agent is the
- 19:54Cloud Code. So for Cloud Code, I create
- 19:57this client credential app. Um there are
- 20:00certain um
- 20:02uh settings that you need to do, the
- 20:04OAuth settings, uh like the callback
- 20:07URL, uh the scopes, the MCP scopes that
- 20:11you need, and uh the the refresh token.
- 20:15Uh and then you will enable the the JWT
- 20:18token access. Um
- 20:21uh and uh yeah, once you do that, you
- 20:23will pretty much be able to like uh
- 20:26request for um the the keys, the
- 20:29consumer key. So you'll get consumer and
- 20:32a key and the secret. And this is very
- 20:34secret for for that specific org um that
- 20:37your clients are going to connect to. So
- 20:39you use that secret and then uh
- 20:42basically authorize this MCP server so
- 20:45you can connect from uh your coding
- 20:48agent. This is what you will going to
- 20:50do. Uh so this is this UI finally loaded
- 20:53here. So as you can see, it has three
- 20:55tools, the Data 360. It has three tools
- 20:58here. This is the server endpoint. And
- 21:01the three tools that we discussed is
- 21:02first to search that right Data 360
- 21:05tool, um and then get the payload
- 21:08example, so it can figure it out, and
- 21:10then it's going to execute that data 360
- 21:12tool.
- 21:13So, having said all of these, right? Um
- 21:17Anush actually built up a skill that we
- 21:20will figure it out, sort of how to like
- 21:22ship this to you, so that you don't have
- 21:24to like click around
- 21:26uh to set it up. It's called set up
- 21:28platform MCP server, basically. And this
- 21:31skill file I I got it from um Anush.
- 21:34Anush, thanks for sharing this.
- 21:36So, really uh Anush, you want to talk
- 21:38about the skill file? I think you built
- 21:39this, yeah.
- 21:40>> Uh yeah, we'll we'll find a way to share
- 21:42that soon. Uh the teams have been
- 21:44working on having this setup skill. Uh
- 21:47we do understand that uh setting up the
- 21:49external client app and connecting it to
- 21:52coding agent currently is a couple of
- 21:55steps.
- 21:55>> Mhm.
- 21:56>> And we want to solve that. So,
- 21:59moving forward soon, it should be as
- 22:01simple as you ask your coding agent set
- 22:04up data 360 using the skill,
- 22:06and it should be able to go set up the
- 22:08MCP server and the external client app
- 22:10for you.
- 22:11>> Yeah. Really whatever I described right
- 22:13now is there in the skill.
- 22:15>> Yeah, there's one caveat though. Make
- 22:17sure that you are the person setting
- 22:19this up as permissions to create an
- 22:21external client app. So, one of the
- 22:22things I've noticed is not everybody in
- 22:24your org might have that permission. So,
- 22:26make sure that you have access to go
- 22:28create the external client app.
- 22:29>> Yeah, perfect. And for those of you who
- 22:31are looking um to like sort of
- 22:35um
- 22:37understand a bit more about like
- 22:40how to like headlessly activate um some
- 22:43of these, right? Um there are
- 22:46uh as I said, there is one command that
- 22:48I can I can share with you right now.
- 22:51It's kind of going uh in uh I would say
- 22:54in in um
- 22:56uh in beta right now,
- 22:58uh kind of beta, uh but you can um sort
- 23:02of ask help about it. It's called SF
- 23:04agent MCP. Um, that's the command. And
- 23:08if you ask for help, it will show you
- 23:11um, everything. Like you can create an
- 23:14MCP server in that API catalog, you can
- 23:16delete it, you can actually fetch uh,
- 23:19MCP server, you can list it, you can
- 23:22update the server. So, this is a command
- 23:24that is also shipping up with the CLI.
- 23:26So, if you're scripting this, um,
- 23:28because I know a lot of lot of you
- 23:30actually work with scratch orgs and
- 23:32sandboxes, and you would want to like
- 23:34just automate this whole thing,
- 23:37um, because the moment you create a
- 23:38scratch org, you want these MCP servers
- 23:40to be activated so that you can connect,
- 23:43uh, easily, right? So, if you're doing
- 23:44that, I suggest go and look at these
- 23:47commands.
- 23:48Um, these are in beta, soon going GA,
- 23:50but this is going to help you in like
- 23:52setting up your scratch orgs or
- 23:54sandboxes from the CI/CD pipeline.
- 23:57Um, okay. So, with with having said
- 24:00that, let's actually move on. And so,
- 24:03once you do that, basically, uh, I have
- 24:06already set up this, um, MCP server. You
- 24:09can see it's already connected in my
- 24:10Cloud Code. I'm using Cloud Code. Um,
- 24:14it's one of the coding agents that I
- 24:15use, but again, this is MCP server. So,
- 24:18you can bring whatever coding agent you
- 24:19like, you know. Some of you I know are
- 24:21using Pi, some of you are using Code X,
- 24:24or you know, some of you are using Agent
- 24:26Force Wipes. We are headless, right? So,
- 24:28the idea with headless is this. Like
- 24:30this whole platform is available to any
- 24:32coding agent. So, but today I'm going to
- 24:35use the Cloud Code because that's what I
- 24:36I use it. So, you can see I'm connected
- 24:38to this Data 360 uh, MCP server. And if
- 24:42you actually go, you can actually view
- 24:44the tools. I'm already authenticated,
- 24:46actually.
- 24:47Um, so, it's all it's all well and good,
- 24:49but if you're not authenticated, there
- 24:51is one,
- 24:52um, command that you will need to type
- 24:54in into your Cloud and provide the
- 24:58the you know, the consumer secret and
- 25:00the consumer key in that command. Make
- 25:03sure you don't again paste the consumer
- 25:04key or the consumer secret into this
- 25:06cloud code tool. Just be on the terminal
- 25:09mode always when you are doing some kind
- 25:11of sensitive um commands, right? So, you
- 25:14you go on a terminal and you do it. Um
- 25:16but once you are connected, right? This
- 25:19tool now can use the MCP server. So,
- 25:22uh I want to show you um
- 25:25uh the data 360 and what I have been
- 25:27building with this scenario to solve the
- 25:29scenario so that we don't do everything
- 25:32here live because we'll run out of the
- 25:33time.
- 25:34So, you can see here um
- 25:36we have this uh diner contact stream.
- 25:40This stream basically is pulling data
- 25:42from Salesforce from contact object. And
- 25:45it maps to a bunch of Again, this whole
- 25:49thing I created um through the MCP
- 25:51server itself. Like I I did not
- 25:54go and do this 100 number of clicks. Um
- 25:57I asked the MCP server to create it. Um
- 26:00and then I said, you know, I want these
- 26:02these mappings for these data and it
- 26:04automatically figured it out like which
- 26:06objects is it should map to.
- 26:09All that Yeah.
- 26:11>> Mohit, before the MCP server, how would
- 26:13you reason through this mappings
- 26:15previously?
- 26:17Just curious.
- 26:18>> Yeah, so one of the things
- 26:21um
- 26:22like one of the things uh how we would
- 26:24reason to this is
- 26:27um
- 26:30So, there are two aspects to to
- 26:32reasoning here. How how um we used to do
- 26:35it.
- 26:36I think there is an auto map feature
- 26:38somewhere in the tool. So, once you
- 26:40click in the UI, it will automatically
- 26:42look at the data types and sort of map.
- 26:47But that will only work until you kind
- 26:49of select that data model entity. Like
- 26:52you will say, you know, select contact
- 26:54and then account contact. So, in this
- 26:56case, and this is a good question from
- 26:58business perspective, how I reason is
- 27:00See, in in data 360, there are only
- 27:03two key things. Whether you want to do
- 27:06the unification and this identity
- 27:09resolution
- 27:10by individual, or you want to do it
- 27:14by the entity, like account, right? And
- 27:18or by business, right? Um so, usually it
- 27:21is mapping to account contact and
- 27:23contact point addresses and point phone
- 27:26and individuals, but in matches by data
- 27:29types. And of course, I have to go by
- 27:31hand and map a lot of these, yeah.
- 27:34>> Yeah, you know, I noticed that at at
- 27:36least for me personally, mapping was a
- 27:38tricky flow because a lot of this had to
- 27:41be hand mapped.
- 27:42>> Mhm.
- 27:42>> Uh so, how was your experience from
- 27:45how you did this previously and now with
- 27:47MCP server?
- 27:49>> Yeah, I would say it was good. All I did
- 27:52was um just ask the the cloud code. I'm
- 27:55going to show one of them. So, this this
- 27:57I built it already for the contact
- 27:59streams and we have one for the app
- 28:00order stream. So, let's do one one from
- 28:03the tool, so everybody can actually see.
- 28:06So, one of the other things as I said,
- 28:07we want to pull is all that review. So,
- 28:09there is a review object. So, typically
- 28:11how you would build is, you know, you
- 28:13will go to new and you know, you will do
- 28:15so many clicks. You will say yes, CRM
- 28:16objects and wait for this screen to load
- 28:20load and then you will be like view the
- 28:22objects. There are too many clicks, you
- 28:24know, too many clicks. My org is I would
- 28:26say reasonable, but still you will you
- 28:28will have to search, etc. etc. find a
- 28:30lot of these things, right? Uh a lot of
- 28:32clicks. Um
- 28:34So, what I'm going to do is I actually
- 28:36wrote a a spec of what I want. Um let's
- 28:40let's look at that spec of the changes I
- 28:42want. So, one of things I'm doing now is
- 28:45this review uh data stream, um
- 28:48basically. So, again, uh
- 28:50again I'm asking what I need to do here
- 28:52in this in this case, right? So, the
- 28:55whole idea here with this spec and you
- 28:58know, all the features that got added in
- 29:00VS code was being able to view this
- 29:02markdown now.
- 29:03So, really what I'm saying is look,
- 29:04contact and app order are ready. I need
- 29:07the review signals ingested into new
- 29:09data stream. And then this is how I like
- 29:12to write my specs actually. Just just
- 29:14So, it is consistent. So, you know,
- 29:16Anush, you it need not be like this. You
- 29:18can just write it however you want. It's
- 29:21It's up to you. But I think it makes it
- 29:23much easier for me also to think, right?
- 29:25In this case I'm saying, you know, once
- 29:27the system will create a data stream for
- 29:29the CRM object
- 29:31and lands as a DLO carrying the rating,
- 29:33review, order, and contact. And then
- 29:36this is the only injection artifact this
- 29:38change creates. And this is built using
- 29:41that MCP discovery loop, right? Not
- 29:44giving again, not giving the payload
- 29:45etc. but just saying it. Scenario is,
- 29:48you know, review DLO is populated and
- 29:50carries the signal. So, this is one
- 29:51scenario. The other scenario is even
- 29:54data stream is provisioned correctly.
- 29:55Like once it's done, make sure it's kind
- 29:57of, you know, provisioned correctly.
- 29:59So, this is basically
- 30:01it. And you know, if you know certain
- 30:03things, like I know there are these many
- 30:06DLOs, these many rows, so I can also
- 30:08say, right? When you query these many
- 30:10rows are there.
- 30:12If you know,
- 30:13sometimes whatever you know, right? That
- 30:16becomes a deterministic signal here.
- 30:18Um
- 30:19So, yeah, simple simple requirement, you
- 30:22know, that I want. Now I'm going to just
- 30:25ask that use this review data stream
- 30:27spec, right? And create a data stream.
- 30:30So, we'll we'll we'll we'll chat about
- 30:32it, but let's submit this work to the
- 30:33agent here. So, we're going to say
- 30:37create
- 30:39a data stream for the reviews.
- 30:43The specs are here.
- 30:47Um yeah, so I'll uh
- 30:51I'll actually submit it right in the
- 30:53auto mode
- 30:54and let the agent kind of figures it out
- 30:58and create that uh stream.
- 31:00Um
- 31:01Yeah, so so Anush, you're you're asking
- 31:03me something. Uh it kind of works
- 31:05through.
- 31:05>> Yeah, I I was cu- I was curious because
- 31:07like I wanted to try to gauge what your
- 31:11workflow was before versus what it is
- 31:13after. Like how have you changed the way
- 31:16in which you operate, right? Like how
- 31:17much of your previous context was
- 31:19important and what can users learn from
- 31:22the way you have changed your uh usage
- 31:24of Data 360?
- 31:26>> Yeah, yeah. So the biggest the biggest
- 31:28unlock for me doing this is see, I'm no
- 31:30more a blocker. You know, we have to
- 31:33think in these terms, right? If I sit on
- 31:35the computer and then like just click
- 31:37around like 50 number of clicks,
- 31:40it's
- 31:42like I I'm I'm the blocker now to to
- 31:44execute that job, right? And if I know
- 31:47that the agent can like, you know,
- 31:49reliably go and do certain aspects of
- 31:51this, right? Uh I'll probably kick that
- 31:54load and then move on, you know, like
- 31:56sort of do something uh else, right? Um
- 31:59so that that has been operational change
- 32:02uh from me is like now I'm spending more
- 32:04time instead of like, you know, clicking
- 32:07around, I'm more spending more time
- 32:09doing the the the design thinking
- 32:13um mostly around, okay, how do I solve
- 32:15this business problem? What are the
- 32:16tools I need? And writing it um in um
- 32:19natural language so I can then queue
- 32:22that work to the agent. And then once
- 32:24the agent kind of does, you know, my
- 32:26work is more of verification. So as you
- 32:28can see, my work is sort of moved moved
- 32:31from like
- 32:32implementing, right? I have more time to
- 32:35do the other side of the work now. Like
- 32:36the the planning, you know, I spend more
- 32:38time doing that planning. After the work
- 32:40I spend more time doing the doing, you
- 32:43know, judgment and all these kind of
- 32:44capabilities. So So that is what I've
- 32:47been doing. Yeah.
- 32:48>> So you moved like you moved up the
- 32:49stack. You have expanded your scope from
- 32:51being a just a developer to being a
- 32:53developer but also like a
- 32:55uh
- 32:55directly responsible individual for
- 32:58coming together with a business plan and
- 33:00evaluating if it is successful or not.
- 33:02>> Absolutely. Absolutely. And you know,
- 33:05this is one terminal that is running as
- 33:07you can see. I My my day now is if if
- 33:10somebody asked me to do something,
- 33:12right? Usually I have like five or six
- 33:13terminals running and I'm just running
- 33:15you know, sort of managing those
- 33:16terminals or those agents in this case
- 33:19we can say, right? Um so you see So so
- 33:22Anush, let's talk about you know, what
- 33:24happened here, right? So you see what
- 33:25the agent did is it said it has the few
- 33:28picture. It knew that the object name is
- 33:31review __c engagement stream. It began
- 33:35that discovery loop, which is basically
- 33:37searching and pulling down the exact
- 33:39schema. So now it has the schema.
- 33:42So now it knows that there is a home
- 33:44connection already in the org. That is
- 33:46so awesome, right? It knows that there
- 33:48is that CRM connection is there and if
- 33:52it would have the connection would have
- 33:53been missing, it will tell me that go
- 33:55first do that connection. Um so I can go
- 33:58go do the connection. So I will be like
- 33:59sort of blocked at this point, right?
- 34:01But now since it has a connection, it's
- 34:02able to go move. So it's now able to go
- 34:05and look at all the things. It looked
- 34:07existing stuff that are there in the in
- 34:09the org. Um it looked into all the
- 34:12prerequisite. It confirmed the the the
- 34:14field and now it's calling the data 360
- 34:18um APIs to I think hopefully this is the
- 34:21next step is to create that um
- 34:23that uh data streams. In fact, I think
- 34:25it's creating. Let's Let's look at see
- 34:28if So you see?
- 34:29There you go. You see it actually
- 34:31created it. So now it's going to
- 34:33schedule the I think it's going to the
- 34:35next thing is it's going to schedule the
- 34:36run, which is to like, you know, get all
- 34:39the
- 34:40the data
- 34:41and schedule that that run.
- 34:44So, it can get all the data into Data
- 34:46360. And
- 34:48it did it so, like, you know, magically,
- 34:50but
- 34:52this this would have been some clicks
- 34:54actually, you know. So, it says, you
- 34:55see, run triggered successfully. Let me
- 34:57poll the DLO account until it populates,
- 35:00right?
- 35:00So, I'm pretty sure it will now go and
- 35:03see, it moved the status to pending. It
- 35:06means it knows that, okay, next I need
- 35:08to kick this job. So, we should see some
- 35:10refresh history. It's not here yet, but
- 35:13it's the next step. So, once it does, I
- 35:16think the next step I'm curious to see
- 35:18is it does the data mapping by itself.
- 35:20If it's not, maybe that's one more
- 35:21prompt. And from there I will have the
- 35:24review data already synced up, sort of
- 35:27without me having to like click, just by
- 35:29intent. I'm I'm moving. Yeah.
- 35:31>> Um meanwhile, I'm also curious,
- 35:34Moeit, uh
- 35:35how did you which model did you pick for
- 35:37this and why did you choose the specific
- 35:39model? Do you have an insight on your
- 35:41usage and what you found works well for
- 35:43you?
- 35:44>> Yeah, yeah, exactly. So, this one the
- 35:47the model that I have picked for this is
- 35:51uh Opus 4.8.
- 35:53Um Opus 4.8
- 35:56is
- 35:57decent model. And the the best thing
- 35:59about Opus 4.8 to me is
- 36:02it somehow magically, like, knows a lot
- 36:05of Data 360 too. You know, I played with
- 36:07this model a lot. The there is that
- 36:09reasoning traces in there, you know,
- 36:11that that understands it does understand
- 36:14the terminology of Data 360 to some
- 36:16extent, right? And again, uh you know,
- 36:19your business might be different and
- 36:21that is the the ontology layer, you
- 36:23know, sometimes you will have to come
- 36:25and write some ontology layer on top of
- 36:29um your
- 36:30you know your work that you already have
- 36:32like lot of you would have already built
- 36:33a lot of these things right if I'm not
- 36:35wrong you you'd have built data streams
- 36:37and data schemas and your you have your
- 36:40own business needs. So as long as you
- 36:42can express that in markdown files or
- 36:46someway to the agent and sort of ground
- 36:48that the agent will like have the
- 36:50context at that point and because it
- 36:53understands the technology of data 360
- 36:55to a lot of extent
- 36:57it will be able to like sort of make or
- 36:59reason through the task really well with
- 37:01these MCP servers of course helping
- 37:03right. One question I have for you
- 37:07Anush is
- 37:08this MCP server right um
- 37:11is it calling the APIs that our
- 37:13developers can also look at somewhere?
- 37:16Is it calling the same APIs?
- 37:17>> Yeah
- 37:19behind behind the scenes the MCP server
- 37:21does call a middleware that does call
- 37:23the same APIs that are exposed to you
- 37:26which is why the
- 37:27>> APIs
- 37:29>> Yeah.
- 37:30>> Okay so so the connect rest APIs right
- 37:32so yeah if you are a developer again we
- 37:34have the developer docs here and you can
- 37:37always go there and look at the connect
- 37:39APIs
- 37:41the ones I'm sure it's actually um
- 37:44hitting right now is I think that the
- 37:46data stream APIs right
- 37:50that is what it is hitting right now
- 37:52but again you see the orchestration of
- 37:54these APIs what needs to happen when um
- 37:59I mean of course for repeatable use
- 38:00cases I would also like if your use case
- 38:03is too complex you know there's no harm
- 38:06in like um you know sort of
- 38:08deterministically writing some of these
- 38:10code or asking an agent to write it and
- 38:12then you run it via the agent. All those
- 38:14kind of workflows are definitely
- 38:16possible but the key point here I want
- 38:18to make out is the agent like sort of
- 38:22loves orchestrating these things. I
- 38:23mean, it has with MCP like those three
- 38:26tools that this team has delivered,
- 38:28right? It knows now these APIs really
- 38:32well and it can reason through that. And
- 38:33the you know, if that makes sense,
- 38:35Anush.
- 38:37Let's see if there are any questions
- 38:38that are coming. I haven't got a chance
- 38:40to look at the the chat here.
- 38:42>> Yeah, that's a good idea.
- 38:44>> Yeah, let's look at
- 38:46um
- 38:47um
- 38:48chat here.
- 38:50Uh is there a way that we can rate limit
- 38:53the calls to these MCP servers?
- 38:56Um Uh so, why is that? So, somebody is
- 38:59asking Himanshu. Um let's see. I'll I'll
- 39:02put the question here so we can all all
- 39:04see here. So, Himanshu is saying is
- 39:07there a way we can rate limit the calls
- 39:08to MCP server APIs in case MCP server
- 39:10connected to the autobots which are used
- 39:12by internal business user and not just
- 39:14coding agent. Thoughts, Anush, on this?
- 39:17>> I believe
- 39:19there is no way to explicitly rate limit
- 39:21the MCP server today.
- 39:23>> Uh-huh.
- 39:24>> On the
- 39:25Salesforce side.
- 39:27>> Mhm.
- 39:27>> But that is we are building governance
- 39:29functionality into that as we go along.
- 39:32But on the client side, while you
- 39:34connect an MCP server to a specific
- 39:36agent, you can always put limits on the
- 39:38client side to figure out, "Hey, do not
- 39:40make more than these number of calls,
- 39:42right?" So, there are client level
- 39:44guardrails. Those are not great yet, I
- 39:46think. But we will be working on having
- 39:49better governance and control for
- 39:51Salesforce admins to have a better
- 39:53organization of how much of an MCP
- 39:56server do you want the users to use.
- 39:58>> Mhm. Mhm. And that also will come with
- 40:01more uh maybe usage metrics and traces
- 40:03and things like that. So, we're still
- 40:04working on some of those governance
- 40:06features.
- 40:07>> Perfect. Perfect.
- 40:09Uh somebody is asking this question,
- 40:11which is really interesting. How does
- 40:13this differ from agent for Salesforce
- 40:15worker?
- 40:16>> So, that's a great question.
- 40:19The MCP server will work with anything
- 40:22that is an MCP client.
- 40:24Now, whether that is Cloud, Codex,
- 40:26GitHub Copilot, whether it's an MCP
- 40:29client, this will work with it,
- 40:30including Agent for Swipes.
- 40:33>> Mhm.
- 40:34>> Agent for Swipes co-worker today is not
- 40:36yet an MCP client.
- 40:38>> I see.
- 40:39>> So, today you will not be able to
- 40:40connect it. But, that doesn't mean in
- 40:42the future you won't be able to. In the
- 40:44future, this could be part of Agent for
- 40:46Swipes co-worker, whether implicitly
- 40:48from our integration or explicitly from
- 40:51your integration.
- 40:52>> Mhm. So,
- 40:53>> uh we're not there yet.
- 40:55Uh the other thing is like
- 40:57you can build your own agent. You don't
- 41:00have to go through any coding agent. If
- 41:01you have your own agent within Pi open
- 41:03code, any MCP client will work with the
- 41:06Data 360 MCP server.
- 41:09>> Mhm. [clears throat] Mhm. Perfect.
- 41:10Perfect. Now, we know this loads takes
- 41:12some time, especially for the first time
- 41:14because it has to bring in all the all
- 41:16the data. Let Let it Let it kind of run,
- 41:19and uh we can um sort of keep moving.
- 41:21So, so really, you know, for this
- 41:23requirement though, my only thing will
- 41:25be build this data stream, right? And
- 41:27then sort of map that to the
- 41:30uh engagement object engagement um
- 41:33type of object category. The other
- 41:35interesting thing to me was when this
- 41:39like when when this data stream got
- 41:40created, it immediately picked up the
- 41:44the category as engagement. Which was
- 41:46interesting. Like, it knew from the
- 41:48business requirement that, you know, the
- 41:50category will not be profile. And this
- 41:52is one of the fundamental decisions that
- 41:54we do in in Data 360 because everything
- 41:57including how we can we are charged,
- 42:00etc. depends on these kind of aspects,
- 42:02like how how are you categorizing the
- 42:04data, right? Um so, it did pick the
- 42:06right stuff, and it's keep on moving. Of
- 42:09course, this is the first it's a data
- 42:10load, so it's going to take some time um
- 42:12behind the scenes running those bad
- 42:14jobs. Oh, we have the data. You see? Oh,
- 42:16no, the fields. Sorry, is the field.
- 42:19Okay, it's going to it's going to take
- 42:20some time,
- 42:21but you can see it's it's active.
- 42:25And it's going to run. The data stream
- 42:27is going to run. And it created the data
- 42:29lake object. So, as I was clicking
- 42:31through, you know, so you see there is a
- 42:32data lake object that got created. So,
- 42:35next would be, you know, DMO creation.
- 42:37And you know, it itself puts like this.
- 42:41I think this is a feature of the harness
- 42:42itself within cloud code. It can like,
- 42:45you know, put itself into the sleep
- 42:46mode. So, it just sleeps and waits
- 42:49for the task completion.
- 42:51So, that's that's perfect. It's it's
- 42:54working it's it's way out. Um
- 42:57Okay, so the other thing that we wanted
- 42:59to sort of discuss here with this
- 43:04is
- 43:05this whole idea of skills, right? Talk
- 43:08us more through that, Anush, like the
- 43:11skills, right? I I see developers more
- 43:13often getting confused like what is
- 43:15skills, what is MCP, do we need MCP
- 43:20at all when we use skills? Um
- 43:23etc. Like, let's let's get to the bottom
- 43:25of that.
- 43:26>> So,
- 43:28I think in the industry there is still
- 43:30some flux going on right now.
- 43:32>> Mhm.
- 43:32>> Um the way we see it is I think there
- 43:35are two differing standards right now
- 43:37coming. You could say one is uh the
- 43:40skills
- 43:42and one is your agent plugins. You might
- 43:44have seen this, right? I think
- 43:46ChatGPT or code it started with agent
- 43:48plugins. A plugin can contain MCP and
- 43:51skills.
- 43:52And then on the other end you have MCPs,
- 43:54which in the next spec will have skills
- 43:56themselves.
- 43:57So, we are still observing that. We will
- 43:59support whichever is the most convenient
- 44:02for our users.
- 44:03>> Mhm.
- 44:04>> So, that is the direction we're seeing
- 44:05things go.
- 44:07>> Okay.
- 44:08>> when we say skills,
- 44:10skills are a way to provide just-in-time
- 44:13context
- 44:14>> Mhm.
- 44:15>> to the agent of your choice.
- 44:17>> I see. I see.
- 44:18>> What do I mean by just-in-time context?
- 44:21That means whenever you have a
- 44:23successful workflow where you do not
- 44:25want the agent to freestyle reason
- 44:28through a problem,
- 44:29>> Mhm.
- 44:30>> but you want it to follow a specific
- 44:32workflow that you know works well,
- 44:34>> Mhm.
- 44:35>> you ask it to invoke a certain skill.
- 44:38>> Got it.
- 44:39>> Now,
- 44:40for Data 360, it could be something like
- 44:44we know best how to create certain
- 44:46segments.
- 44:47>> Mhm.
- 44:48>> We know how to do transforms.
- 44:50>> Right.
- 44:50>> If you ask an agent to do transforms, an
- 44:53agent might reason through all the world
- 44:55knowledge it has
- 44:56>> Mhm.
- 44:57>> that is relevant and irrelevant to Data
- 44:59360.
- 45:00>> Got it.
- 45:01>> And then it will consume tokens. It will
- 45:04take a lot of tokens.
- 45:06It might still get it right,
- 45:08but then you're spending a lot of tokens
- 45:09doing it.
- 45:10>> Right. Right.
- 45:11>> And it becomes slightly more
- 45:13non-deterministic.
- 45:14>> Mhm.
- 45:15>> So
- 45:16Yeah. So when the
- 45:20expert in a certain product, now that
- 45:22could be the Salesforce product
- 45:24developers or engineers
- 45:26>> Mhm.
- 45:26>> or the Salesforce implementers in the
- 45:28community, when you know how best to
- 45:31implement a certain thing and you do not
- 45:33want the agent to do some freestyle
- 45:35reasoning,
- 45:37codifying your workflow in the form of a
- 45:39skill and then invoking that will
- 45:41provide great results.
- 45:43>> I see. Okay.
- 45:45>> One of the trend I've seen is also that
- 45:48if you pick a really good model with
- 45:50high reasoning, it can actually do most
- 45:53things without skills.
- 45:54>> Right.
- 45:55>> But then that becomes expensive,
- 45:56consumes tokens.
- 45:58Now the question for me is how do you
- 46:00take like a Claude Sonnet or Haiku model
- 46:04that is really low cost and make it
- 46:06perform at the same level of Opus.
- 46:08>> Right.
- 46:09>> So,
- 46:10cost is one place where skills are
- 46:11helpful, speed is another place where
- 46:13skills are helpful, and having more
- 46:15deterministic execution is where skills
- 46:17are helpful.
- 46:18>> Got it.
- 46:19>> Got it.
- 46:19>> Yeah, that was enlightening. That was
- 46:21enlightening. Now, while you are there,
- 46:23you know, one of the things is you see
- 46:25it actually pulled the record. So, that
- 46:27job was success.
- 46:29So, it it has a refresh history now. So,
- 46:32that job is successful now. So, it's
- 46:35moving on to the next thing. So, now it
- 46:37says the this is all thing, you know,
- 46:39these are scheduled synchronously. It's
- 46:42able to infer the conclusion also,
- 46:44right? Um
- 46:47So, uh batch is still pending. Now,
- 46:50at this time, you know, because it is
- 46:52like in the sleep mode, you see, I can
- 46:54see in the UI that it's all successful.
- 46:56But, because the agent sort of put
- 46:57everything in into the sleep mode, um
- 47:00it's it's going to sleep over for like
- 47:02these these many time. And then it's
- 47:04going to go back and query and uh
- 47:07uh sort of tell that, "Okay, this is
- 47:09this is successful. Let's let's move on,
- 47:11right?"
- 47:12Um what I'm going to do is
- 47:14of course, stop it because we know it's
- 47:17successful. I'm going to say it's
- 47:19successful right now.
- 47:22Um
- 47:23and uh can you move ahead?
- 47:27Mapping, creating DMOs, and mapping it,
- 47:31right?
- 47:32Um
- 47:33Now, again, this is because I know the
- 47:36next step, so I'm able to tell. But, if
- 47:38you leave it, it is automatically going
- 47:40to take that step, you know, because
- 47:42again, it has the reasoning like this
- 47:45this Claude knows that after this, this
- 47:47is the step that needs to happen, right?
- 47:49Um
- 47:50so, again, it's going to go and work
- 47:52this out. It should
- 47:54Now, it should be very simple because
- 47:55now the next step it's going to do is,
- 47:57you know of course map it right like if
- 47:59I if I do I will need to
- 48:01pull that DMO but I don't think I have
- 48:03the DMO here.
- 48:05Let's see.
- 48:07You see it's going to every click takes
- 48:09some time right in data 360 because it's
- 48:12it has to search through all the objects
- 48:13is a big system.
- 48:15Um so
- 48:17I don't think I have the diner review
- 48:19object it.
- 48:22You see I don't have the review object
- 48:24okay we don't have the review object it.
- 48:25So the agent is going to create it. So
- 48:28now it's calling you see it's call it's
- 48:29called four times.
- 48:31Uh let me confirm the review count then
- 48:33move on right up in a knot to build a
- 48:35unified DMO it's to use this mapping on
- 48:37the contact to auto provision and
- 48:39standard and then map to the other
- 48:40sources.
- 48:41Now it's going to it's going to do. So
- 48:43this is a
- 48:45I think it learned from somewhere the
- 48:46reasoning here for harmonization. Um
- 48:49that it always needs to look at what is
- 48:51there. So it's the the first reasoning
- 48:53it did is like yeah I need to be able to
- 48:55map using the standard mappings.
- 48:58But I'm pretty sure it will try to find
- 49:00out and then it says no there's no way
- 49:02to map this review object to contact
- 49:04because it's not same. It's going to
- 49:06reason like that. Um it's very different
- 49:08than how I think actually
- 49:11uh but you can see that it actually
- 49:13comes to the same conclusion maybe doing
- 49:15three or four more turns on it.
- 49:18Um
- 49:19but it's it's it's it's different. So
- 49:22you can see it will now create the
- 49:24mappings. It will take some time I'm not
- 49:25going to do it manually. The the agent
- 49:27is going to create the mapping. So once
- 49:29it's create the mapping I'll ask it to
- 49:31create the
- 49:32the identity resolution. Um again uh
- 49:35just for all of you right I will want to
- 49:37show really like what what would be my
- 49:40next step here or what is it that we're
- 49:42going to do here. So you can see the
- 49:44review you we have the reviews object we
- 49:46have the contact team and we have the
- 49:47app order object right so we have all of
- 49:49these. So the next step is to actually
- 49:52use all of this and sort of map it to
- 49:55individual objects within
- 49:57um
- 49:58within Salesforce and sort of relate all
- 50:00of these
- 50:01um
- 50:02and then uh from there, the idea is to
- 50:05create this uh identity resolution job,
- 50:07this one.
- 50:09Now, I just need to wait uh for um the
- 50:12DMOs to be created here. So, standard
- 50:14mapping succeeded. Now, I need to map
- 50:17the app orders DLO. So, now it's
- 50:19mapping. So, you see the mappings is
- 50:20happening now. So, once it maps, my next
- 50:23step would be again, as I said, create
- 50:25the
- 50:26So, you see the SSO individual is
- 50:27populated. Now, let's get the app order
- 50:30DLO.
- 50:31And yeah, next my next intent will be to
- 50:33create that identity resolution job and
- 50:35it's going to go go create it. Now, I
- 50:38know a lot of you would have done
- 50:39identity resolution and you know the
- 50:41pain. It's like you have to click at
- 50:44least 10 to 12 or maybe more clicks in
- 50:47here to actually do certain things in
- 50:50Data 360. And I'm so glad that this MCP
- 50:53server is here and it's able to like
- 50:56you know, sort of automate all of this
- 50:58and uh instead of me sitting through and
- 51:00clicking
- 51:01um the system can like reason through.
- 51:04Uh let's see if there are any any
- 51:06questions that are coming.
- 51:09Uh
- 51:11using headless 360, how to pass data
- 51:13effectively react as a front-end
- 51:16framework?
- 51:17Uh very simple, actually. Um this is
- 51:20something I've I've myself done it. It's
- 51:23everything in Data 360 can be queryable
- 51:26through Apex. So, Apex recently, and let
- 51:29me show you guys this one.
- 51:31Um Apex connect query.
- 51:34Um
- 51:36So, Apex um
- 51:38uh connect query for Data 360. Apex
- 51:41actually launched uh a new modern
- 51:45um query class, actually, that you can
- 51:48use it to just query yeah, it's called
- 51:51SFQL query namespace. This is a new
- 51:54namespace actually. Um, I've used it
- 51:57often. So, once you know your tables,
- 52:00you're just writing SQL statements
- 52:01within Apex.
- 52:03Basically very simple. Try it out and
- 52:05then you know, you can build react apps
- 52:07or whatever on Salesforce. And then just
- 52:09use these to query um, from data 360. We
- 52:12we made it much more easier now to just
- 52:15natively sort of use it. Uh, let's see
- 52:18if there are any any other questions
- 52:19here.
- 52:20Um, this is one. How is your diner data
- 52:24stream able to function when um, fields
- 52:28have not been mapped? You might be in
- 52:30the process of configuring it.
- 52:32Um, yeah. So, the diner data stream So,
- 52:35first you create the data stream and of
- 52:37course it's not mapped, but
- 52:39um, you know, yeah, we need the mapping
- 52:41before we do the identity resolution.
- 52:42That's why I'm waiting for it. Let's see
- 52:44if it is actually done.
- 52:47Uh, let's see where the agent is.
- 52:50Um, yeah, agent is working through it.
- 52:52You can see
- 52:54uh,
- 52:56most of the effort creating. So, it is
- 52:59So, the other thing that Anusha is saw
- 53:01this is one of the wonderful things is
- 53:03actually goes back and looks what is
- 53:05there in the org.
- 53:06>> Yeah.
- 53:07>> Um,
- 53:08>> That itself was previously taking me a
- 53:11lot of time, right? Like whenever I need
- 53:12to go to context switch between what I'm
- 53:14doing and go back to the org and go back
- 53:17to my documentation, it's saving me a
- 53:19lot of that time recently.
- 53:21>> Yeah, and I think why it is doing is
- 53:23because I have this skill called reuse
- 53:25before reinvent. So, this is one of the
- 53:27skills I wrote it in my harness. And my
- 53:31plan is very simple is like anytime the
- 53:33agent builds something, I wanted to go
- 53:36and see if there is anything it can
- 53:38reuse.
- 53:39It can backfire also sometimes. Like it
- 53:42will try to It will try to use um things
- 53:46that it shouldn't. Um but I have the
- 53:49skill. So, So, again, uh this is very
- 53:51interesting thing, right, with the the
- 53:53skills concept here.
- 53:54>> Yeah.
- 53:55>> Like, yeah, it all depends on you. Like,
- 53:57the hardness
- 53:58I mean, there is a thing like there will
- 54:00be a a baseline hardness that somebody
- 54:03can provide. But, then how you work um
- 54:07if you want to like
- 54:09like keep that non-shareable, then you
- 54:12can do that as well. Like, you can
- 54:14create your own skills and then use it
- 54:16within your hardness. So, that's that's
- 54:18very powerful
- 54:19um capability. So, yeah, it's calling
- 54:23It's calling It's It's doing some some
- 54:25stuff here. Let's see.
- 54:28Yeah, it's it's doing It's basically
- 54:30going to create that object and then I
- 54:32think, yeah, the identity resolution, I
- 54:34think it will
- 54:36it will create. I I need to ask it to
- 54:38create identity resolution. Currently,
- 54:40it is actually listing.
- 54:41So, it's asking me. See, the standard
- 54:44DMOs. So, it kind of asking me the
- 54:46standard DMOs and the contact order
- 54:48mapping already exist. What should I
- 54:50build next? So, create that DLO. Yeah.
- 54:55Yeah. Skip further mapping and build the
- 54:58IR. Of course, we can't build the IR
- 55:00without we know that we have to map the
- 55:02the DLOs onto the individual with the
- 55:04party and the rating. Uh this is the
- 55:07missing DMO. Yeah. So, I'm going to go
- 55:09with that option.
- 55:10This is that loop. This is what I call
- 55:13it loop. If you are upfront like
- 55:17doing all the decisioning
- 55:19you will never get in caught into these
- 55:21kind of loops, right? We can get rid of
- 55:23these kind of loops where it ask your
- 55:24feedback, but sometimes you can design
- 55:26loops where, you know, you are also part
- 55:29of the loop. So, yeah. So, now it now
- 55:32it's going to build that DMO. I think
- 55:34the next step is that. And if I leave my
- 55:37agent, it will next build identity
- 55:39resolution, segmentation, whatnot. In
- 55:41fact, you know, I had a spec that I I
- 55:44ran through this. Uh let me show you
- 55:46that spec.
- 55:48It's called Diner Winback. And uh just
- 55:51because, you know, this whole thing can
- 55:53take like hours of long-running task, we
- 55:56we are short of time here. But, this is
- 55:59the spec I wanted to show that I
- 56:01actually ran in this org. So, I
- 56:04indicated all the source system. I
- 56:06exactly told what I need to do.
- 56:09Um I told it how to segment it, um how
- 56:12to unify it, uh and then my scope of the
- 56:15work and the test boundaries, and my
- 56:17acceptance criteria and my assumptions.
- 56:19And uh you know, some things like, you
- 56:22know, I wanted to
- 56:24like keep a a companion doc of how to
- 56:27build like my own version of it. I also
- 56:29wanted it to inventory what is being
- 56:31built out so I can review it. This is
- 56:33another thing. If you're using a
- 56:35long-running agent, it will it can go
- 56:37and build like tons of components, and
- 56:39then you will have to reason out when
- 56:41you're judging it. So, it is always a
- 56:43better idea to inventory it um so that
- 56:46you can go and verify stuff. Because,
- 56:48look, verification is something that
- 56:50agent
- 56:51cannot do. Like, it it can do in some
- 56:53way, but it still needs like, you know,
- 56:55your your judgment here. And that is
- 56:57very critical now than ever.
- 56:59>> uh I I'm curious, like, do you want to
- 57:01take a minute to explain what is your
- 57:04process to create the spec? And what's
- 57:06your process to evaluate the what the
- 57:08agent builds? Because, sometimes we do
- 57:11feel when building is so easy, it can
- 57:13feel slightly lazy to go and uh verify
- 57:16the work. I do feel that sometimes.
- 57:18Yeah. And manual verification does not
- 57:20scale at the rate in which agents can
- 57:23build things, right? So, what's your
- 57:24process to evaluate uh whether the right
- 57:26thing has been done or not?
- 57:28>> Yeah, that's a good question. So, here's
- 57:31the thing. I I I still think for
- 57:33enterprise applications, right?
- 57:36Verification is definitely
- 57:39needs to happen at human level. So,
- 57:43especially for enterprise because
- 57:44there's a lot of things at stake. Now,
- 57:46this is also a judgement to begin with
- 57:48like if you know that this thing hasn't
- 57:52that it will not have that many impact,
- 57:54then you can move really faster, right?
- 57:56Like sometimes I'm not even reviewing
- 57:58the code. I'm like, "Yeah, okay. What
- 57:59about the agent did? Let's just test it
- 58:01out." Like smoke test, it's working
- 58:03fine.
- 58:04But sometimes
- 58:06if it's very critical, right?
- 58:08It depends on the degree of criticality
- 58:10to answer in short, right? If it's very
- 58:11critical, I I generally ask agent to
- 58:15like do some queries and you know,
- 58:17automate that way like you know, query
- 58:19it through this data lake objects. And
- 58:22if I have a smoke test data given, I
- 58:25will sort of manually do or at least
- 58:29manually keep an inventory that, "Okay,
- 58:31if this thing happens, this is
- 58:34you know, this is what I'm expecting,
- 58:35right?" And then I ask agent to go and
- 58:37run that loop for me. So, it will go and
- 58:39run that loop, right?
- 58:41So, I can trust what the agent did is
- 58:44exactly correct or not.
- 58:46>> Yeah.
- 58:46>> Uh
- 58:47but somebody like
- 58:49it's still somebody has to keep the
- 58:51system
- 58:53in their head like entire head like
- 58:55because see, it's like if you let agent
- 58:57sort of take over that thing from you,
- 59:00it will go and build a lot of things
- 59:01that you have no idea and that's the
- 59:03worst thing. Like whatever agent build
- 59:06and if you don't have an idea about it,
- 59:08that is the worst problem to have,
- 59:10right? So, that's that's the shift that
- 59:12is happening like because this
- 59:14implementation is so these agents can
- 59:16implement stuff very faster, right?
- 59:18>> Yep. Yeah.
- 59:19>> Um the the other part of the works
- 59:21suddenly become so much more important,
- 59:23right?
- 59:24>> Yes.
- 59:24>> Um yeah. So, so that's the thing
- 59:27um that I would say. And uh yeah, I
- 59:30think the agent is still kind of working
- 59:32through it. So look what it did. So
- 59:36here's the remaining new net mapping.
- 59:38Let me get the the field it's trying to
- 59:40call. But I don't know I expect a
- 59:43possibly category I'll execute it.
- 59:45Cannot map engagement. You see it tried
- 59:48to map to the wrong thing and it
- 59:51suddenly realized it is not the right
- 59:53thing to do. Now remember like you know
- 59:55at this point that is the human signal
- 59:57is so much more important because for
- 59:59the next iteration I can put this in the
- 1:00:01skill. In the skill file. Like whenever
- 1:00:04you are mapping these type of objects go
- 1:00:07map to this, go map to that. So now you
- 1:00:10know instead of like your point right?
- 1:00:12It will go back and sort of
- 1:00:15figure it out somehow.
- 1:00:17These agents are goal driven. You have
- 1:00:19given it a goal. They're going to figure
- 1:00:21it out one way or the other. They're
- 1:00:23going to reason keep reasoning. You know
- 1:00:25like as you said you know it might have
- 1:00:26wasted ton of tokens here doing some
- 1:00:29unnecessary work but it is coming to the
- 1:00:30same conclusion here. See it was like
- 1:00:33contact point email mapping was created
- 1:00:35it did create it did create some
- 1:00:37mappings here. It's moving on.
- 1:00:39But my point is same like see it created
- 1:00:42the mappings now. You see
- 1:00:44it went and you know
- 1:00:46figured out that it needs to map to the
- 1:00:48contact point email. So we have a common
- 1:00:51email now, right? Now it's going to
- 1:00:53create another object. Because it needs
- 1:00:56that review object and it's going to map
- 1:00:58the other fields to that. And then once
- 1:01:00it's ready we'll go create the identity
- 1:01:02resolution rule. So as you said like you
- 1:01:04know sometimes the recipe is in your
- 1:01:06head and you want to just communicate to
- 1:01:08the agent. Sometimes you just let the
- 1:01:10agent loose let it figure it out you
- 1:01:11know and both of these worlds have their
- 1:01:13own you know advantages
- 1:01:16and their own drawbacks as you said like
- 1:01:18drawbacks and pros and I think that is
- 1:01:20where the human judgment suddenly
- 1:01:22becomes like so important, right?
- 1:01:25Uh yeah, so it's going to work out. I'm
- 1:01:27going to leave my laptop, um you know,
- 1:01:29and uh as you can see, like these things
- 1:01:32can take for ever, like, you know, a a a
- 1:01:35couple of hours. One of the settings I
- 1:01:38did with my laptop is not to sleep, you
- 1:01:40know. This is the first change that
- 1:01:41happened after these agents and I saw
- 1:01:44how they work, right?
- 1:01:45>> Yep.
- 1:01:45>> So, they can work and I think that is
- 1:01:47where the cloud loads are also going to
- 1:01:49become more critical, like, you can push
- 1:01:51this workload to the cloud and let all
- 1:01:53these things, you know, sort of happen
- 1:01:55in the cloud. I can see that uh thing
- 1:01:58happening as well. Uh in fact, not
- 1:02:00happening. It's uh it's already
- 1:02:01happening. Uh with that, actually, I
- 1:02:03want to move back to the slides,
- 1:02:04actually. I mean, this is going to
- 1:02:06happen. Um I did run, as I said, with
- 1:02:09that spec and it it built end-to-end
- 1:02:11everything, which was surprising. Um
- 1:02:14anyway, so with that, right? Uh I think
- 1:02:16we are running short of the time. I
- 1:02:18don't want to take much of your time as
- 1:02:19well, Anush, but for all of you, right,
- 1:02:21joining today, a big thank you, right?
- 1:02:24Thank you for joining us and I hope you
- 1:02:26learned and saw this MCP. Now, one of
- 1:02:29the other things I want is feedback.
- 1:02:31Definitely, that feedback is there. Um
- 1:02:34take a screenshot of this one. I'm going
- 1:02:35to keep it for a minute here. Um
- 1:02:38this is all the resources that we have.
- 1:02:40I'll also try to do a trailblazer
- 1:02:42community post sharing all these things.
- 1:02:45Um here we have blog post where we
- 1:02:47introduce this DX360 MCP server. Um
- 1:02:51there is another code live sessions and
- 1:02:54um
- 1:02:55uh that is not the code live session. I
- 1:02:58think this session This is on-demand
- 1:03:00video that we have to learn about code
- 1:03:02extensions. I think they are also great.
- 1:03:04You can ask agent to write code to do
- 1:03:06certain aspects, like, rack pipelines or
- 1:03:09if you're building um transformations.
- 1:03:12Um and then, um you know, if you want to
- 1:03:15learn about cloud code, how to build um
- 1:03:18agent skills with Cloud Code, there is a
- 1:03:20link to there where we talk about that
- 1:03:23aspect of how to use Cloud Code to sort
- 1:03:25of build your own own skills. So, all of
- 1:03:28these resources are there. And then,
- 1:03:30yeah, of course, um you know, the other
- 1:03:32other aspect here I want you to remind
- 1:03:34is um the Data Blazer community, right?
- 1:03:38Um please post all your questions.
- 1:03:40Anusha's there as well.
- 1:03:42>> Yeah, I'll be there to answer if you got
- 1:03:44If you guys have any questions from this
- 1:03:46session, you want some follow-ups, you
- 1:03:48want resources that you have not gotten
- 1:03:50yet, me and Moit as well as our
- 1:03:53engineering team will be in the Data
- 1:03:55Blazer community, and we'll be able to
- 1:03:57help you out. You can start a post about
- 1:03:59today's session, ask any questions, and
- 1:04:02we'll do our best to make you
- 1:04:03successful.
- 1:04:05>> Awesome. Awesome.
- 1:04:06Yeah, with that, I think um you know, we
- 1:04:09are end of this. And again, uh you know,
- 1:04:11for all of you, um I I hope this was
- 1:04:14educational. And uh I hope you're going
- 1:04:17to play with the MCP. Please let us know
- 1:04:20if there is anything else we can help.
- 1:04:22Any parting thoughts here uh from you,
- 1:04:23Anush?
- 1:04:24>> Uh no, this has been great, Moit. I
- 1:04:26think uh there's nothing better to use
- 1:04:29or learn about the MCP server than a
- 1:04:31hands-on session, right? This is not
- 1:04:33something you want to hear people just
- 1:04:34talking about it, but actually showing
- 1:04:36how it works. And
- 1:04:38I think you've done a great job in
- 1:04:40educating all of us in what your
- 1:04:42workflow was before, how it's changed
- 1:04:44now, and the way things are moving. So,
- 1:04:47uh I'm really glad. Thank you for
- 1:04:48everyone joining you, and thank you for
- 1:04:50Moit as well for conducting this session
- 1:04:52for us.
- 1:04:53>> Yeah, absolutely. Absolutely. And uh
- 1:04:56yeah, um with with that, right? Um
- 1:04:59again, a lot of you are asking in the
- 1:05:00chat about how can we share some of
- 1:05:03these um skills and open-source work
- 1:05:06that as open-source work, um
- 1:05:09we'll definitely make all the possible
- 1:05:11attempts here to give you the
- 1:05:12documentation and the skills. Um and
- 1:05:16also the MCP server though is live, you
- 1:05:18know, make use of it. Connect it to your
- 1:05:20coding agent if you're working in data
- 1:05:22360 space.
- 1:05:24Uh your developer edition all or also
- 1:05:26has all of these access. So, if you're
- 1:05:28just learning, please make use of it. Um
- 1:05:31and yeah, um you know, um
- 1:05:33we will make sure that we will give you
- 1:05:35all the all the resources
- 1:05:37uh and always
- 1:05:38please feel free to reach out to us.
- 1:05:41And with that,
- 1:05:42thank you. Thank you everyone.
- 1:05:43>> All right.
About this transcript
This page contains the full transcript of Agentforce NOW codeLive: Build Agentic Data 360 Experiences Headlessly by Salesforce Developers, generated from the public captions YouTube serves with the video. The transcript has 11,682 words across 1,787 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.