YT YouTube2Text

YouTube2Text → Transcripts → Agentforce NOW codeLive: Build Agentic Data 360 Experiences Headlessly

Agentforce NOW codeLive: Build Agentic Data 360 Experiences Headlessly — Transcript

by Salesforce Developers · 11,946 words · 1,747 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:03[music]
0:10[music]
0:12Hey,
0:17hey, [music]
0:22hey. [music]
0:34>> [music]
0:41>> Hey. [music]
0:46Hey. Hey. [music]
0:51Heat. Heat. N.
0:53[music]
1:00[music]
1:08>> [music]
1:15>> Hey, [music]
1:20hey,
1:21[music]
1:24hey.
1:33Doo
1:39[music]
1:48be
1:53down.
1:58Hey,
2:04[music]
2:11[music]
2:18[music] hey, hey.
2:21Heat. Heat.
2:24[music]
2:37[music]
2:42[music]
2:46Heat. Heat. N.
2:52[music]
3:02[music]
3:06[music]
3:12>> [music]
3:18[music]
3:26[music]
3:31[music]
3:39[music]
3:47[music]
3:56[music]
4:34[music]
4:43>> Heat. Hey, Heat.
4:52Heat. Heat. N.
4:56[music]
5:02Heat. Heat. N. [music]
5:12Heat. Heat. N.
5:22Hey, hey, hey.
5:26[music]
5:42Heat.
5:46[music]
5:55[music]
5:59Heat.
6:06[music]
6:23>> [music]
6:29[music]
6:32>> Heat.
6:38[music]
6:43[music]
6:48[music]
6:50Hey, Heat.
6:53[music]
6:59>> [music]
7:05[music]
7:09[music]
7:14>> Hey,
7:16[music]
7:17hey, hey.
7:21Heat. Heat.
7:24[music]
7:28[music]
7:35[music]
7:44[music]
7:52>> [music]
7:56>> Heat.
7:59[music] Heat.
8:06[music]
8:10>> [music]
8:16>> Hey.
8:19[music]
8:28[music]
8:34[music]
8:36Heat.
8:39[music]
8:46[music]
8:51[music]
8:56[music]
8:58Hey, heat. Hey, heat.
9:02Always
9:05[music]
9:29[music]
9:37>> [music]
9:45[music]
9:50>> Oh yeah. [music]
9:57Hey. Hey.
9:59Heat. Heat.
10:02[music]
10:07[music]
10:21[music]
10:27>> [music]
10:35[music]
10:42[music]
10:49[music]
10:58[music]
11:06[music]
11:09>> down.
11:12[music]
11:13Hello.
11:17>> [music]
11:25[music]
11:30[music]
11:34>> Hey. [music]
11:42[music]
11:46Heat. Heat.
11:50[music]
11:59[music]
12:10>> [music]
12:16[music]
12:26[music]
12:28>> Hey,
12:32hey, hey.
12:35Heat. Heat. [music]
12:46[music]
12:56[music]
13:04>> [music]
13:11[music]
13:18[music]
13:33>> Hey,
13:37[music]
13:46hey, [music]
13:47hey.
14:08Hey,
14:12hey, hey.
14:17Heat.
14:19[music]
14:33[music]
14:36Heat.
14:52>> [music]
14:55>> All right. All right. Hello everyone.
14:58Thank you so much for joining Agent
15:00Force now code live. My name is Mohit.
15:03I'm a developer advocate here at
15:06Salesforce and I welcome you all to this
15:08code live episode. Today we're going to
15:10be building agentic data experiences
15:15more importantly headlessly all through
15:17coding agents. Uh and uh I have with me
15:21uh the product manager who is sort of
15:23working on all of this Anush. Welcome to
15:27code live.
15:28>> Yeah thank you Mo. Thank you for
15:29inviting me. Uh hi everyone. My name is
15:32Anush. I'm the product manager for
15:34headless data 360 starting with the data
15:37360 MCP server. So really excited to for
15:40this hands-on session with Moeth to get
15:42you guys started and to see your
15:44experience using the MCP server.
15:46>> Perfect. Perfect. And and I know Anush
15:48you're very busy building these MCP
15:50servers. So I really appreciate you
15:53taking time today and for all those of
15:56you who are tuning um remember that we
15:58also have chats there. So as we sort of
16:01show these capabilities to you, if you
16:04have any questions, please make use of
16:06those chat and u shoot us to us and
16:10we'll we'll get back to those questions
16:12as they sort of come along. uh with that
16:15uh you know because this is Salesforce
16:17presentation um I'll I'll put the
16:19presentations here but more importantly
16:22uh a forward-looking statement uh so
16:25please uh do not make your purchasing
16:27decisions based on features
16:30um that are uh not available in the
16:32product at or not GA with that out of
16:35our way again um you know myself and
16:38Anush we're going to be talking about
16:40building agentic data 360 experiences
16:42headlessly Okay. And I just want to
16:45remind everyone that uh you know there
16:47is a QR code. I've also uh sort of
16:50placed this QR code on your screens on
16:52your right. So um make sure that if you
16:55have any feedback about this code live
16:58or any code live in general just uh use
17:01that QR code and tell us with that. Um
17:05how about uh Anush uh we actually take
17:07some time and explain uh to everyone um
17:12like what is this data 360 headless all
17:14about? What is this momentum about? What
17:16are these MCP servers? So let's take
17:18some time everyone to get to the same
17:20page here.
17:21>> Yes. Uh I think that'd be a good start.
17:24Uh starting with what are we doing here?
17:26If you started earlier this year using
17:28data 360, maybe even in Jan, uh a lot of
17:32the functionality was in the lightning
17:34UI in Salesforce. So even if you're
17:37using agents back then, you had to move
17:39between the UI, maybe ask your coding
17:42agent what you need to do, look at the
17:44documentation. So this led to like some
17:46sort of context mismatch between where
17:48you're doing work and where you need to
17:50go gather this context for. So in order
17:53to solve that we are starting a really
17:56focused headless program.
17:59What does this mean?
18:01>> The foundation of this program for us is
18:04making sure all our functionality in the
18:06UI is available through APIs. So that is
18:08the foundation layer in which we are
18:11building. Once those APIs are exposed we
18:15have now released a data 360 MCP server.
18:18Actually earlier in June we had a
18:20developer preview of a local MCP server
18:23right through which we got a lot of
18:24feedback about how users are using it.
18:27We saw the user experience with a lot of
18:29great feedback from that and from that
18:32we realized that the hosted remote MCP
18:34server was a better option for most
18:36users.
18:38>> Perfect.
18:39>> Now
18:41why is this really important? Right?
18:44It's not just about you going to claude
18:46or going to Codex and telling uh Codex
18:49go do this thing or claude do this thing
18:50no make no mistakes. It's not just that.
18:53It's about providing trusted context
18:56also with governed resources and
19:00pre-built skills via the MCP server. Why
19:03is this important?
19:06So if you notice, agents are good at
19:10solving things, but they may also not be
19:13the best at following the shortest path
19:15to a solution. But you being Salesforce
19:18developers or industry experts know what
19:21is the best path. So this MCP server
19:24paired with specialized skills reduce
19:27the number of tokens required for an
19:29agent to go from 0 to one. So if you
19:33notice the MCP server today, we have
19:35just three tools. We have the search
19:37tool, we have payload examples and we
19:41have execute.
19:43Why just three tools, right?
19:46We notice that with dynamic tool loading
19:49when you ask for a certain intent, the
19:53agent can then source for what what it
19:55needs. For example, it could be
19:56segmentation. you want to create a user
19:59segment to maybe uh target a specific
20:02user segment and then the agent finds
20:05the exact payload examples for that tool
20:07and that reduces the number of turns an
20:10agent needs to successfully solve this.
20:13So which is why we have followed this
20:14dynamic tool loading method as well as
20:17coming soon will be skills accompanied
20:20with the MCP server. This will reduce
20:22the number of turns your agent needs,
20:24the number of tokens you need and also
20:26it provides best practices that
20:28practitioners using the data 360
20:31platform use to build it.
20:35So now not just can you do this through
20:37cloud or codeex but you get this really
20:39smart reasoning engine paired with you
20:42following the best practices and
20:44executing on it which is why I'm so
20:46excited for this.
20:47>> Awesome. Awesome. No, this is this is
20:49great. Um one thing I really want to
20:51call out is look at the simplicity of
20:54this whole design right uh when I when I
20:57was sort of um thinking about you know
21:00these MCPS for data 360 and myself you
21:04know um um I I can call myself data 360
21:08uh developer or you know data blazer uh
21:13so there are there are so many tools in
21:16data 360 as as you know um Anush Like
21:20there's so much of power that this data
21:22360 has and the tools it has you know
21:25let's talk about you know creating these
21:27data streams to bring data from like you
21:29know multiple system like thousands of
21:31systems right you can ingest data
21:33through various mechanisms like for
21:36example zero copy is one of my favorite
21:39because if you have a snowflake or data
21:41bricks you can you know sort of zero
21:42copy data into data 360 uh but still
21:47like all of those involve lot of clicks
21:49And when I thought about MCPS for you
21:51know data 360 with all these tools I was
21:55like this is going to be like you know
21:57maybe n number of tools right thousands
21:59of tools here in this in this scenario
22:01because there are thousands of APIs
22:03>> but uh you know when I saw like okay you
22:06know you guys are like only exposing
22:08like only three different MCP tools um
22:12so that's again less number of token
22:14consumptions and also much easier for
22:17the AI to like sort of reason right
22:19because it's now progressively sort of
22:21progressing towards what what is
22:24required than you know just loading
22:26everything in one shot and sort of
22:28consuming the context window and then of
22:30having AI just hallucinate it right
22:34>> perfect perfect okay uh so for all of
22:36you right um today what what we are
22:39thinking is we're going to make this
22:41much more like practical just showing
22:43the use of it um and uh if you're new to
22:47data 360 welcome um no issues we we'll
22:50try to like cover those aspects as well
22:53like what is data 360 what does it do
22:55etc um but I have a business use case
22:58and I think you know all of us are there
23:01to solve business problems right so I
23:03thought it will resonate well if we
23:05start with the the use case here and uh
23:08Anush for us today the use case is we
23:11have this bronto a food delivery
23:12platform and this is a sample
23:14application that developer advocacy team
23:16sort of uses it uh to showcase uh
23:19different capabilities of the platform.
23:22So we've taken that Pronto uh platform
23:24today and we will leverage the data 60
23:27as a centralized system for context for
23:30agents that Pronto is building. Um
23:34basically it's going to be unifying data
23:36from multiple sources
23:38activating that data across different
23:40applications and different agents. So
23:42this is the primary use case of data 360
23:45in case you are wondering what is data
23:46360 right is that context engine that
23:49will drive agents uh your apps you know
23:52think of it like a big data processing
23:54engine right with all the tools and
23:56batteries included so you can quickly
23:58activate you know the data um for
24:01context as well as for um sort of
24:05building applications if if that's what
24:07you're doing with this data right um so
24:10that's so our use case our exad business
24:12problem that we're going to solve. We're
24:14going to try to solve in this session.
24:15Although I don't know how far we will
24:17get along with these coding agents and
24:19sort of working because you know
24:21building something live is like you know
24:23hours of work like even clicking through
24:25it and agents is going to simplify a lot
24:27of it. We'll try to to sort of do a lot
24:29of stuff here um live. So really what we
24:32are trying to solve is um for pronto the
24:36customer is showing up now in as two
24:38different peoples even though it's same
24:40customer um but it's showing up as two
24:43different people because uh you know one
24:47way the information is coming through
24:48their delivery app system which uses uh
24:52basically an application which uses the
24:54the email to order and then they are
24:57collecting all the reviews in Salesforce
25:00CRM And we need to a able to sort of
25:04activate it using data from multiple of
25:07these systems. The rules that we're
25:09going to write is not just belongs to
25:11the reviews here like our rules here as
25:13you can see is two different rules here.
25:16One is if the order date is sort of
25:18lapsed um 60 days beyond and also the
25:22rating or the rating is less than two.
25:24So you you can see we are creating a
25:26segmentation operation to sort of win
25:29back these customers from multiple
25:31systems and these are exact type of
25:33problems actually uh at scale that data
25:36360 sort of solves right
25:37>> yeah so this is a common problem I've
25:39seen moit where a lot of customers may
25:42have unhappy customers they're serving
25:45>> and how do you how do you figure out
25:46that's the right customer right when you
25:48have this data among multiple different
25:49systems how do you know this reviewer is
25:52actual customer of you is that a real
25:54review. Is it a fake review? How do you
25:56figure out should you reach out to them
25:57or is it worth your time?
25:59>> Perfect. Perfect. Yeah. Yeah. No, this
26:00is the use case. In fact, you used to
26:02call the CDP use case. I I remember the
26:05term CDP use case, right? Um but here's
26:09here's the simple design to solve these
26:11kind of things and I just put it out
26:13here because this is very common um you
26:17know solution design that you um that
26:20you as solution engineers are are doing
26:22most of the times. So you will basically
26:24in data 360 create data streams uh for
26:29different data um sources. In this case
26:32it is that app the data coming from the
26:35app uh which in my case I've simplified
26:38it actually with just a CSV file but
26:40think about it's coming from your mobile
26:42app or something. And then um uh the d
26:46uh the other data uh source here is the
26:48Salesforce itself that has all the
26:50reviews. So we take all those. So we
26:52create data streams. We create um what
26:55we call uh data lake objects uh and then
26:59uh from that we create uh data model
27:02objects by mapping them to the you know
27:04the canonical data model that um data
27:07360 provides. So data 360 by default has
27:11pre-built data models for different
27:14business purposes to activate that data
27:17uh or unify that data. Right? So, so
27:20that is unification. So, we'll do that
27:21unification generally that that is your
27:23workflow creating streams and doing this
27:26unification or mapping it to the right
27:29canonical data model within the data 360
27:32and then you resolve using identity
27:34resolution. In this case, it's very
27:36simple like I've just taken email as
27:38that common thing but it could be you
27:40know multiple logic like multiple stuff
27:42uh it could be very complex too uh fuzzy
27:45matches etc. So based on that you map
27:47them and match it and create that one
27:50single identity record from which we can
27:53create this segmentation which will be
27:55another uh data object called winback
27:58diners uh you know where we sort of
28:01segment all those customers who have
28:04like sort of not ordered from pronto app
28:07uh uh you know it's been like 60 days
28:09since they ordered or u the rating has
28:13like sort of fallen less than um True.
28:16So this is the solution design. Now
28:20uh you know because I have worked with
28:21this system I I know this solution
28:23design. uh but when you're solving a
28:26problem right and because we have
28:28different tool sets another way to
28:30approach this kind of problem these days
28:33with the agents is
28:36ask the agent itself like how do I solve
28:40using that MCP that we have and the
28:43agent will definitely sort of look at
28:44all the payloads and can recommend that
28:46these are the things again in this loop
28:49you still need to be in the loop because
28:51you need to judge whether whatever that
28:53agent is sort of telling you is correct
28:55or not. So you still need to understand
28:57all of these u but uh once you sort of
29:00understand then this whole MCP uh and
29:05skills and everything becomes much more
29:07useful because then there are like you
29:08know power tools which you can put them
29:10into action and can move with a very u
29:14you know rapid pace. Um any thoughts
29:16here um Anush? Yeah. So, uh I think you
29:20really nailed that one because even for
29:23me when I if for me sometimes the
29:25business problem is just I want to win
29:27back these dinos like I do not know
29:29anything about data 360.
29:31>> I don't know I need to unify this data.
29:32All I need to do is my goal is I need to
29:34win back these customers who have
29:36dropped off from the platform, right?
29:38And
29:39>> I usually pair with Claude or Codeex and
29:41I'm like, "Hey, this is my problem. This
29:43is what I have at my disposal. Help me
29:45brainstorm through the solution." And as
29:48long as you're connected to the tools to
29:50the agent, whether it's cloud, cursor,
29:52codeex, or even GitHub copilot, they'll
29:55be able to help you reason through your
29:57various options. And uh like Moit
29:59mentioned since Moit's a data 360 expert
30:02he has a good grasp of what he needs to
30:05do that is unify resolve and segment but
30:07if you are new to data 360 just ask your
30:10coding agent hey this is my problem how
30:13can I solve this and you'll be able to
30:14brainstorm through it
30:16>> yeah yeah exactly perfect um so let's
30:20let's go into that live coding segment
30:22with a coding agent um and uh the data
30:25uh 360 MCP so I want to take this MCP
30:28server for a ride today and see how it
30:30kind of does. Um and again um you know
30:33we want to be upfront honest with you.
30:35So this is all GA right. Um there are
30:38certain things in here that we are still
30:40bridging like of course like data 360
30:42has a lot of like lot of APIs and lot of
30:46capabilities right and to bring all of
30:48them to the MCP I think uh Anush's uh
30:51Anush and the team is working on but I
30:53would say from my experience working on
30:55this now from like last two weeks I can
30:57say you know definitely this is this is
31:00u helping me like you know I'm able to
31:02um sort of move with the speed of my
31:05intent here um you know just knowing
31:07what I need to do and then uh using this
31:10MCP is has been a helper. Anush you have
31:12any thoughts on what are the things that
31:14you're still working on. Um so yeah so
31:17anything that's a UI functionality u it
31:21will be exposed to the API and you will
31:23see it in the MCP server.
31:25>> If there are some recent APIs you should
31:27be seeing them shortly in the MCP
31:29server. But there are also certain APIs
31:32or certain functionality we might not
31:34expose.
31:35>> Mhm. An example for that is uh imagine
31:38you need to set up a connection which
31:39requires putting your credentials into
31:42an API or credentials into the UI to
31:44connect from system to system,
31:46>> right?
31:46>> We might not want you to put that those
31:48credentials into the agent because then
31:51that might be stored somewhere might
31:53which might lead to a security lapse.
31:56>> Yeah. So you might notice very few
31:59sliver of functionality that we might
32:02intentionally keep away from the lens of
32:07security as well as governance.
32:09>> But whatever other functionality that
32:11you have, you will see it in there.
32:13>> If we if you have missed something,
32:15please reach out to us, let us know. If
32:17something doesn't work out for you,
32:18well, let us know and we'll be we'll
32:20make sure to address that as soon as
32:21possible.
32:22>> Yeah. Perfect. Uh perfect. Um so first
32:25actually so I wanted to show you the
32:27Pronto instance. So this is the the
32:29Pronto instance that I was talking about
32:31and uh this is basically what I'll be
32:34using for my uh building uh demos. So
32:36one of the things we will do is ask the
32:39agent to create data streams um and then
32:42create identity resolutions today with
32:45um data streams from Salesforce um
32:48pulling all the reviews. So these are in
32:50Salesforce. Now we want to pull it into
32:51data data 360 here so we can unify with
32:54the app orders data. So that's something
32:56we'll be doing. But before all this
32:59there is some setup that is involved in
33:03um sort of setting up your org so that
33:05that MCP gets activated and you can
33:08connect to your coding agent. I want to
33:10go through that now uh with you. So uh
33:14real quick right um you should find
33:17this MCP server in your API catalog. So
33:22if you go to API catalog you will see
33:24something like data 360 server. So um
33:28again for every org you will basically
33:32go there um and this is going to take
33:35some some time. Let me actually refresh.
33:37Sometimes it does take some time to load
33:39but really what it will have is it'll
33:42have a toggle. Now there is already a
33:45Salesforce CLI command to that the CLI
33:49team is exposing. It's in beta at this
33:51point through which you can um sort of
33:54not have to like go through the UI to
33:57activate the servers. You can just like
34:00do it through the CLI itself. So that
34:03way you can script it if you need to.
34:05Uh, let's see why it's kind of hanging
34:07up here. Let's give it a shot once more.
34:12This is why I like headless actually.
34:17The headless makes it so easy. I don't
34:19have to click around and sort of wait
34:21for, you know, uh, this UI to sort of
34:23load up. Uh, but while that is loading
34:26up, right, uh, once that is sort of
34:29activated, which it will, uh, it will
34:31show up here, but let me show you what
34:34you need to do. So, so the first step is
34:35obviously to activate it. The next step
34:38here is to um create something called as
34:42external uh client credential app and uh
34:47it's in my case my coding agent is the
34:51cloud code. So for cloud code I create
34:54this client credential app. Um there are
34:57certain um uh settings that you need to
35:01do the oath settings uh like the call
35:04back URL uh the scopes the MCP scopes
35:08that you need and uh the the refresh
35:11token uh and then you will enable the
35:14the JWT token access. Um uh and uh yeah
35:19once you do that you will pretty much be
35:22able to like uh request for um the the
35:26keys the consumer key. So you'll get
35:28consumer and key and the secret and this
35:31is very secret for for that specific org
35:34um that your clients are going to
35:35connect to. So you use that secret and
35:38then uh basically authorize this MCP
35:42server so you can connect from uh your
35:45coding agent. This is what you going to
35:47do. Uh so this is this UI finally loaded
35:50here. So as you can see it has three
35:52tools. The data 360 it has three tools
35:55here. This is the server endpoint and
35:58the [snorts] three tools that we
35:59discussed is first to search that right
36:02data 360 tool um and then get the
36:05payload example so it can figure it out
36:07and then it's going to execute that data
36:09360 tool. So having said all of these
36:12right um Anush actually built up a skill
36:16that we will figure it out sort of how
36:18to like ship this to you so that you
36:21don't have to like click around uh to
36:23set it up. It's called setup platform
36:26MCP server basically and this skill file
36:29I I got it from um Anush. Anush thanks
36:32for sharing this.
36:33>> So really uh Anush you want to talk
36:34about the skill file? I think you this.
36:37>> Yeah. Uh yeah, we'll we'll find a way to
36:39share that soon. Uh the teams have been
36:41working on having this setup skill. Uh
36:44we do understand that uh setting up the
36:46external client app and connecting it to
36:49coding agent currently is a couple of
36:52steps.
36:52>> Mhm.
36:53>> And we want to solve that. So moving
36:56forward soon it should be as simple as
36:58you ask your coding agent set up data
37:01360 using the skill and it should be
37:04able to go set up the MCP server and the
37:06external client app for you.
37:08>> Yeah, pretty much whatever I described
37:10right now is there in the skill.
37:12>> Yeah, there's one caveat though. Make
37:14sure that you are the person setting
37:16this up as permissions to create an
37:18external client app. So one of the
37:19things I've noticed is not everybody in
37:21your org might have that permission. So
37:23make sure that you have access to go
37:25create the external client app.
37:26>> Yeah. Perfect. And for those of you who
37:28are looking um to like sort of
37:32um
37:34understand a bit more about like how to
37:37like headlessly activate
37:40um some of these right um there are uh
37:44as I said there is one command that I
37:46can I can share with you right now. It's
37:48kind of going uh in uh I would say in in
37:52u uh in beta right now uh kind of beta
37:57uh but you can um sort of ask help about
38:00it. It's called SF agent MCP. Um that's
38:04the command and if you ask for help it
38:07will show you um everything like you can
38:10create an MCP server in that API
38:12catalog. You can delete it. You can
38:14actually fetch a MCP server. you can
38:17list it, you can update the server. So
38:20this is a command that is also shipping
38:22up with the CLI. So if you are scripting
38:24this um because I know a lot of lot of
38:26you actually work with scratch arcs and
38:29sandboxes and you would want to like
38:31just automate this whole thing um
38:34because the moment you create a scratch
38:35arg you want these MCP servers to be
38:38activated so that you can connect uh
38:40easily right so if you're doing that I
38:42suggest go and look at these commands um
38:45these are in beta soon going GA but this
38:48is going to help you in like setting up
38:49your scratch arcs or sandboxes from the
38:52CI/CD pipeline. Um, okay. So with with
38:56having said that, let's actually move
38:58on. And so once you do that, basically,
39:02uh, I have already set up this, um, MCP
39:05server. You can see it's already
39:07connected in my cloud code. I'm using
39:09cloud code. Um, it's one of the coding
39:12agents that I use. But again, this is
39:14MCP server. So you can bring whatever
39:16coding agent you like. You know some of
39:18you I know are using pi some of you are
39:20using codeex
39:22you know some of you are using agent
39:23force wipes we are headless right so the
39:26idea with headless is this like this
39:27whole platform is available to any
39:29coding agent so but today I'm going to
39:32use the cloud code because that's what I
39:33I use it so you can see I'm connected to
39:36this data 360 uh MCP server and if you
39:39actually go you can actually view the
39:41tools I'm already authenticated actually
39:44um so it's all it's all well and good
39:46But if you're not authenticated, there
39:48is one um command that you will need to
39:51type in into your cloud and provide the
39:55the you know the consumer secret and the
39:58consumer key in that command. Make sure
40:00you don't again paste the consumer key
40:01or the consumer secret into this cloud
40:04code tool. Just be on the terminal mode
40:06always uh when you are doing some kind
40:08of sensitive um commands. Right? So you
40:11you go on a terminal and do it. Um but
40:14once you are connected right this tool
40:16now can use the MCP server. So uh I want
40:20to show you um uh the data 360 and what
40:24I have been building with this scenario
40:26to solve this scenario so that we don't
40:28do everything here live because we'll
40:30run out of the time. So you can see here
40:32um we have this uh diner contact stream.
40:36This stream basically is pulling data
40:39from Salesforce from contact object and
40:42it maps to a bunch of again this whole
40:46thing I created um through the MCP
40:48server itself like I I did not go and do
40:52this 100 number of clicks um I asked my
40:54MCP server to create it um and then I
40:58said you know I want these these
41:00mappings for these data and it
41:01automatically figured it out like which
41:03objects [clears throat] it should map to
41:06all that option. Yeah.
41:08>> Moit before the MCP server, how would
41:10you reason through these mappings
41:12previously? Just curious.
41:15>> Yeah. So, one of the things um like one
41:19of the things uh how we would reason
41:21through this is
41:24um
41:27so there are two aspects to to reasoning
41:30here. How how uh we used to do it. I
41:33think there is an auto map feature
41:35somewhere in the tool. So once you click
41:37in the UI, it will automatically look at
41:40the data types and sort of map. But that
41:44will only work until you kind of select
41:47that data model entity. Okay? Like you
41:49will say you know select contact and
41:51then account contact. So in this case
41:54and this is a good question from
41:55business perspective. How I reason is
41:57see in in data 360 there are only two
42:00key things whether you want to do the
42:04unification and this identity resolution
42:07by individual or you want to do it by
42:12the entity like account right or by
42:15business right um so usually it is
42:18mapping to account contact and contact
42:21point addresses and point phone and
42:24individuals but it matches by data types
42:26and uh of course I have to go by hand
42:29and map a lot of these. Yeah.
42:31>> Yeah. Even I noticed that at at least
42:33for me personally mapping was a tricky
42:36flow because a lot of this had to be
42:38handmapped.
42:39>> Uh so how was your experience from how
42:42you did this previously and now with MCP
42:45server?
42:46>> Yeah, I would say uh it was good. All I
42:49did was um just ask the the cloud code.
42:52I'm going to show one of them. So this
42:54this I built it already for the contact
42:56streams and we have one for the app
42:57order stream. So let's do one one from
43:00the tool so everybody can actually see.
43:03So one of the other things as I said we
43:04want to pull is all that review. So
43:06there is a review object. So typically
43:08how you would build is you know you'll
43:10go to new and you know you will do so
43:12many clicks. You will say yeah CRM
43:14objects and wait for this screen to load
43:16load and then you will be like view the
43:19objects. There are too many clicks you
43:21know too many clicks. my org is I would
43:23say reasonable but still you will you'll
43:26have to search etc etc find a lot of
43:28these things right a lot of clicks um so
43:31what I'm going to do is I actually wrote
43:33a spec of what I want um let's let's
43:37look at that spec of the changes I want
43:40so one of the things I'm doing now is
43:42this review uh data stream um basically
43:45so again um again I'm asking what I need
43:49to do here in this in this case Right.
43:52So the whole idea here with this spec
43:54and you know one of the features that
43:56got added in VS code was being able to
43:58view this markdown now. So really what
44:01I'm saying is look contact and app order
44:03are ready. I need the review signals
44:05ingested into new data stream and then
44:08this is how I like to write my specs
44:10actually um just just so it is
44:12consistent. Um so you know Anush you it
44:14need not be like this. You can just
44:16write it however you want. It's it's up
44:18to you but I think it makes it much
44:20easier for me also to think right in
44:22this case I'm saying you know once the
44:24system will create a data stream for the
44:26CRM object um and lands as a DLO
44:30carrying the rating review order and
44:31contact and then uh this is the only
44:34injection artifact this change creates
44:37uh and this is built using that MCP
44:39discovery loop right not giving again
44:41not giving the payload etc but just
44:43telling it scenario is you know review
44:46DLO is populated and carries the signal.
44:48So this is one scenario. The other
44:50scenario is event date stream is
44:51provisioned correctly like once it's
44:53done make sure it's kind of you know
44:55provisioned correctly. So this is
44:57basically um it and you know if you know
45:00certain things like I know there are
45:02these many dlos these many rows so I can
45:05also say right when you query these many
45:07rows are there uh if you know um
45:10sometimes whatever you know right that
45:13becomes a deterministic signal here. Um
45:16so yeah simple simple requirement uh you
45:19know um that I want and now I'm going to
45:21just ask that use this review data
45:24stream uh spec right and create a data
45:26stream. So we'll we'll we'll we'll chat
45:29about it but let's submit this work to
45:30the agent um here. So, we're going to
45:33say create um a data stream for the
45:37reviews.
45:40The specs are here.
45:44Um yeah. So, I'll uh
45:48I'll actually submit it right in the
45:50auto mode and let the agent kind of
45:53figures it out and create that uh
45:56stream. Um yeah. Yeah. So, so Anush,
46:00you're you're asking me something uh
46:02kind of works through.
46:02>> Yeah. I I was I was curious because like
46:05I wanted to try to gauge what your
46:08workflow was before versus what it is
46:10after like how have you changed the way
46:13in which you operate, right? Like how
46:15much of your previous context was
46:16important and what can users learn from
46:19the way you have changed your uh usage
46:21of data 360?
46:23>> Yeah. Yeah. So the biggest the biggest
46:25unlock for me doing this is see I am no
46:28more a blocker you know we have to think
46:30in this terms right if I sit on the
46:32computer and then like just click around
46:35like 50 number of clicks it's like I I'm
46:40I'm the blocker now to to execute that
46:42job right and if I know that the agent
46:45can like you know reliably go and do
46:47certain aspects of this right uh I'll
46:50probably kick that load and then move on
46:52you know like sort of do something uh
46:55else right um so that that has been
46:58operational change uh from me is like
47:00now I'm spending more time instead of
47:02like you know clicking around I'm more
47:05spending more time doing the the the
47:08design thinking um mostly around okay
47:11how do I solve this business problem
47:13what are the tools I need and writing it
47:16um in um natural language so I can then
47:19cue that work to the agent and then once
47:21the agent kind of does you know my work
47:23is more of verification. So as you can
47:26see my work is sort of moved moved from
47:28like implementing right I have more time
47:32to do the other side of the work now
47:33like the the planning you know I spend
47:35more time doing that planning
47:37>> after the work I spend more time doing
47:39the doing you know judgment and all
47:41these kind of capabilities so so that is
47:43what I've been doing. Yeah. So you moved
47:45like you moved up the stack. You have
47:47expanded your scope from being a just a
47:49developer to being a developer but also
47:51like a directly responsible individual
47:54for coming together with a business plan
47:57and evaluating if it was successful or
47:59not.
48:00>> Absolutely. Absolutely. And you know
48:02this is one terminal that is running as
48:04you can see I my my day now is if if
48:07somebody asks me to do something right
48:09usually I have like five or six
48:10terminals running and I'm just running
48:12you know sort of managing those
48:14terminals or those agents in this case
48:16we can say right um so you see so so
48:19Anush let's talk about you know what
48:21happened here right so you see what the
48:22agent did is it said it has the f
48:25picture it knew that the object name is
48:28review
48:30c engagement stream. It began that
48:32discovery loop which is basically
48:34searching and pulling down the exat
48:36schema. So now it has the schema. So now
48:40it knows that there is a home connection
48:42already in the org. That is so awesome,
48:44right? It knows that there is that CRM
48:47connection is there and if it would have
48:49the connection would have been missing,
48:51it will tell me that go first do that
48:53connection. Um so I can go do the
48:55connection. So I will be like sort of
48:56blocked at this point, right? But now
48:58since it has the connection it's able to
49:00go move. So it's now able to go and look
49:02at all the things. It looked existing
49:05stuff that are there in the in the org.
49:08>> Um it looked into all the prerequisite.
49:10It confirmed the the the field and now
49:13it's calling the data 360 um APIs to I
49:17think hopefully this is the next step.
49:18It's to create that um that uh data
49:21streams. In fact I think it's creating.
49:23Let's let's look at see. So you see
49:26there you go. You see it actually
49:28created it. So now it's going to
49:30schedule the I think it's going to the
49:32next thing is it's going to schedule the
49:33run which is to like you know get all
49:36the the data uh and schedule that that
49:40run um so it can get all the data into
49:43data 360 and it did it so like you know
49:47magically but um this this would have
49:50been some clicks actually you know. So
49:52it says you see run triggered
49:54successfully. Let me pull the DL count
49:56until it populates. Right? So I'm pretty
49:58sure it will now go and see it moved the
50:01status to pending. That means it knows
50:04that okay next I need to kick this job.
50:06So we should see some refresh history.
50:08It's not here yet but it's the next
50:11step. So once it does I think the next
50:13step I'm curious to see is it does the
50:15data mapping by itself. If it not, maybe
50:18that's one more prompt and from there I
50:20will have the review data already synced
50:23up um sort of without me having to like
50:25click just by intent I'm I'm moving.
50:27Yeah.
50:28>> Um meanwhile I'm also curious Mo uh how
50:33did you which model did you pick for
50:34this and why did you choose the specific
50:36model or do you have an insight on your
50:38usage and what you found works well for
50:40you?
50:41>> Yeah. Yeah. Exactly. So this one the the
50:45model that I have picked for this is uh
50:48Opus 4.8.
50:50Um Opus 4.8
50:53is decent model and the the best thing
50:56about Opus 4.8 to me is it somehow
51:00magically like knows a lot of data 3602.
51:03You know I played with this model a lot.
51:05The there is that reasoning traces in
51:08there. you know that that understands it
51:11does understand the terminology of data
51:13360 to some extent right and again uh
51:16you know your business might be
51:17different and that is the the ontology
51:20layer you know sometimes you'll have to
51:21come and write some uh ontology layer on
51:25top of um your um you know your work
51:28that you already have like a lot of you
51:29would have already built a lot of these
51:31things right if I'm not wrong you you'd
51:33have built data streams and data schemas
51:35and you you have your own business
51:38needs. So as long as you can express
51:40that in markdown files or
51:43some way to the agent and sort of ground
51:45that the agent will like have the
51:47context at that point and because it
51:50understands the technology of data 360
51:52to a lot of extent um it will be able to
51:55like sort of make or reason through the
51:57task really well with these MCP servers
52:00of course helping right. One question I
52:02have for you um Anush is this MCP server
52:07right? Um
52:08is it calling the APIs that our
52:11developers can also look at somewhere?
52:13Is it calling the same APIs?
52:15>> Yeah. Uh behind behind the scenes the
52:18MCP server does call a middleware that
52:20does call the same APIs that are exposed
52:22to you
52:23>> which is why the
52:24>> these API
52:26>> Yeah.
52:27>> Okay. So so the connect rest API right?
52:29So yeah, if you're a developer, again,
52:31we have the developer talks here and you
52:34can always go there and look at the
52:36connect APIs. Um the ones I'm sure it's
52:40actually um hitting right now is I think
52:43the the data stream uh APIs, right? Uh
52:47that is what it is hitting right now. uh
52:49but again you see the orchestration of
52:51these APIs what needs to happen when um
52:56I mean of course for repeatable use
52:58cases I would also like if your use case
53:00is too complex you know there's no harm
53:03in like um you know sort of
53:05deterministically writing some of these
53:07code or asking an agent to write it and
53:09then you run it via the agent all those
53:12kind of workflows are definitely
53:13possible but the key point here I want
53:16to make out is the agent like sort
53:19loves orchestrating these things. I mean
53:21it has with MCP like those three tools
53:24that this team has delivered right it
53:26knows now these APIs really well and it
53:29can reason through that and that you
53:31know if that makes sense Anush
53:33>> let's see if there are any questions
53:35that are coming I haven't got a chance
53:37to look at the the chat here
53:39>> yeah that's a good idea
53:41>> yeah let's look at um some chat here uh
53:47is there a way that we can rate limit
53:50the calls to these MCP servers. Um, so
53:54why is that? So, somebody is asking
53:56Himmanshu. Um, let's see. I'll I'll put
53:59the question here so we can all all see
54:01here. So, Himmanu is saying, is there a
54:04way we can rate limit the calls to MCP
54:06server APIs in case MCP connected to the
54:08autobots which are used by internal
54:10business user and not just coding agent?
54:13Thoughts Anush on this? I believe there
54:16is no way to explicitly rate limit the
54:18MCP server today
54:21>> on the Salesforce side
54:24>> but that we are building governance
54:26functionality into that as we go along
54:29but on the client side while you connect
54:31an MCP server to a specific agent you
54:34can always put limits on the client side
54:36to figure out hey do not make more than
54:38these number of calls right so there are
54:40client level guardrails those are not
54:42great yet I But we will be working on
54:45having better governance and control for
54:48Salesforce admins to have a better
54:50organization of how much of your MCP
54:53server do you want the users to use.
54:55>> Mhm. Mhm.
54:56>> And that also will come with more uh
54:59maybe usage metrics and traces and
55:01things like that. So we're still working
55:02on some of those governance features.
55:04>> Perfect. Perfect. Uh somebody is asking
55:07this question which is really
55:09interesting. How does this differ from
55:11agent force coworker?
55:13So that's a great question.
55:16The MCP server will work with anything
55:19that is an MCP client. Now whether that
55:22is Claude, Codeex, GitHub, Copilot,
55:25whether it's an MCP client, this will
55:27work with it, including agent FS wipes.
55:30>> Mhm.
55:31>> Agent F's coworker today is not yet an
55:34MCP client.
55:35>> I see.
55:36>> So today you will not be able to connect
55:38it, but that doesn't mean in the future
55:40you won't be able to. In the future,
55:41this could be part of agent folds
55:43coworker whether implicitly from our
55:46integration or explicitly from your
55:48integration.
55:49>> So, uh we are not there yet.
55:52>> Uh the other thing is like
55:54you can build your own agent. Don't have
55:57to go through any coding agent. If you
55:58have your own agent within PI, open
56:00code, any MCP client will work with the
56:03data 360 MCP server.
56:06>> Mhm. Mhm. Perfect. Perfect. Now we know
56:09these loads take some time especially
56:10for the first time because it has to
56:12bring in all the all the data. Let let
56:14it let it kind of run and uh we can um
56:17sort of keep moving. So so really you
56:19know for this requirement though my only
56:21thing will be build this data stream
56:23right and then sort of map that to the
56:27uh engagement object engagement um type
56:30of object category. The other
56:32interesting thing to me was when this
56:36like when when this data stream got
56:38created, it immediately picked up the
56:40the category as engagement which was
56:43interesting like it knew from the
56:45business requirement that you know the
56:47category will not be profiled and this
56:49is one of the fundamental decisions that
56:51we do in in data 360 because everything
56:54including how we we are charged etc
56:57depends on these kind of aspects like
56:59how how are you categorizing the data
57:02Right. Um so it did pick the right stuff
57:04and it's keep on moving. Of course this
57:06is the first it's a data load so it's
57:08going to take some time um behind the
57:10scenes running those bad jobs. Oh we
57:12have the data you see. Oh no the fields.
57:14Sorry is the field. Okay it's going to
57:16it's going to take some time u but you
57:19can see it's it's active
57:22and uh it's going to run the data stream
57:24is going to run and it created the data
57:26lake object. So as I was clicking
57:28through you know so you see there is a
57:29data lake object that got created. So
57:32next would be you know DMO creation and
57:35uh you know it itself puts like this I
57:38think this is a feature of the harness
57:39itself within cloud code. It can like
57:42you know put itself into the sleep mode.
57:44So it just sleeps and waits uh for the
57:47task completion. Um so that's that's
57:50perfect. It's it's working. It's it's
57:52way out. Um okay so the other thing that
57:56we wanted to um sort of discuss here
57:59with this um is
58:02this whole idea of skills right talk
58:05talk us more through that um Anush like
58:08the skills right I I see developers more
58:10often getting confused like what is
58:12skills what is MCP do we need MCP
58:16uh at all when we use skills um etc like
58:21let's let's get to the bottom of that.
58:24>> So, uh I think in the industry there is
58:27still some flux going on right now.
58:29>> Mhm.
58:30>> U the way we see it is I think there are
58:32two deferring standards right now
58:34coming. Uh you could say one is uh the
58:37skills and one is your agent plugins. Uh
58:41you might have seen this right? I think
58:43chat GPD or codec started with agent
58:45plugins. A plug-in can contain MCP and
58:48skills. And then on the other end you
58:50have MCPs which in the next spec will
58:52have skills themselves. So we are still
58:55observing that we will support whichever
58:58is the most convenient for our users.
59:01>> So that is the direction we're seeing
59:03things go.
59:04>> Okay.
59:04>> So today when we say skills
59:07skills are a way to provide just in time
59:11context
59:12>> to the agent of your choice.
59:14>> I see. I see.
59:15>> What do I mean by just in time context?
59:18That means whenever you have a
59:20successful workflow where you do not
59:22want the agent to freestyle reason
59:25through a problem
59:27>> but you want it to follow a specific
59:29workflow that you know works well.
59:32>> You you ask it to invoke a certain
59:34skill.
59:35>> Got it.
59:36>> Now for data 360 it could be something
59:40like we know best how to create certain
59:43segments.
59:45>> We know how to do transforms.
59:47>> Right? If you ask an agent to do
59:49transforms, an agent might reason
59:51through all the world knowledge it has.
59:54>> Mhm.
59:54>> That is relevant and irrelevant to data
59:56360.
59:57>> Got it.
59:58>> And then it will consume tokens. It will
1:00:01take a lot of turns. It might still get
1:00:04it right. But then you're spending a lot
1:00:06of tokens doing it.
1:00:07>> Right. Right.
1:00:08>> And it becomes slightly more
1:00:10nondeterministic.
1:00:12>> So
1:00:13>> yeah. So when the expert in a certain
1:00:18product now that could be the Salesforce
1:00:21product developers or engineers
1:00:23>> or the Salesforce implementers in the
1:00:26community when you know how best to
1:00:28implement a certain thing
1:00:30>> and you do not want the agent to do some
1:00:32freestyle reasoning
1:00:34codifying your workflow in the form of a
1:00:36skill and then invoking that will
1:00:38provide greater results.
1:00:40>> I see. Okay.
1:00:42One of the trend I've seen is also that
1:00:45if you pick a really good model with
1:00:47high reasoning, it can actually do most
1:00:50things without skills,
1:00:51>> right?
1:00:52>> But then that becomes expensive,
1:00:53consumes tokens. Now the question for me
1:00:56is how do you take like a clot sonnet or
1:00:59a haiku model that is really low cost
1:01:02and make it perform at the same level of
1:01:04opus.
1:01:05>> Right?
1:01:06>> So cost is one place where skills are
1:01:08helpful. Speed is another place where
1:01:10skills are helpful and having more
1:01:12deterministic execution is where skills
1:01:14are helpful.
1:01:15>> Got it. Got it. Yeah, that was
1:01:17enlightening. That was enlightening.
1:01:19Now, while you are there, you know, one
1:01:20of the things is you see it actually
1:01:22pulled the record. So, that job was
1:01:25success. So, it it has a refresh history
1:01:29now. So, that job is successful now. So,
1:01:31it's moving on to the next thing. So,
1:01:34now it says the this is all thing you
1:01:36know these are scheduled synchronously.
1:01:39it's able to infer the conclusion also
1:01:41right um so uh batch is still pending
1:01:46now at this time you know because it is
1:01:49like in the sleep mode you see I can see
1:01:51in the UI that it's all successful but
1:01:53because the agent sort of put everything
1:01:55in into the sleep mode um it's it's
1:01:58going to sleep over for like these these
1:02:00many time and then it's going to go back
1:02:02and query and uh uh sort of tell that
1:02:05okay this is this is successful let's
1:02:07let's move on. Right. Um what I'm going
1:02:10to do is of course stop it because we
1:02:14know it's successful. I'm going to say
1:02:16it's successful right now.
1:02:19Um
1:02:21and uh can you move ahead
1:02:24mapping creating DMOS and mapping it
1:02:28right? Um now again this is because I
1:02:32know the next steps so I'm able to tell
1:02:35but if you leave it it is automatically
1:02:37going to take that step you know because
1:02:39again it has the reasoning like this
1:02:42this claude knows that after this this
1:02:44is the step that needs to happen right
1:02:46um so again it's going to go and work
1:02:49this out it should now it should be very
1:02:52simple because now the next step it's
1:02:53going to do is you know of course map it
1:02:55right like if I if I do I will need to
1:02:58pull that DM DMO, but I I don't think I
1:03:00have the DMO here. Let's see.
1:03:04You see, it's going to every click takes
1:03:06some time, right, in data 360 because
1:03:08it's it has to search through all the
1:03:10objects. It's a big system. Um, so I
1:03:14don't think I have the diner review
1:03:16object yet.
1:03:19You see, I don't have the review object.
1:03:21Okay, we don't have the review object.
1:03:23So, the agent is going to create it. So,
1:03:25now it's calling. You see, it's it's
1:03:26called four times. uh let me confirm the
1:03:29review count then move on write up in
1:03:31order not to build a unified DMO it's to
1:03:33use this mapping on the content to
1:03:35autoprovision and standard and then map
1:03:36to the other sources now it's going to
1:03:39it's going to do so this is
1:03:42I think it learned from somewhere the
1:03:43reasoning here for harmonization
1:03:46um that it always needs to look at what
1:03:48is there so it's the the first reasoning
1:03:50it did is like yeah I need to be able to
1:03:52map using the standard mappings but I'm
1:03:55pretty sure it will try to find out and
1:03:57then it says no there's no way to map
1:03:59this review object to contact because
1:04:01it's not same it's going to reason like
1:04:04that um it's very different than how I
1:04:06think actually uh but you can see that
1:04:09it actually comes to the same conclusion
1:04:12maybe doing three or four more turns on
1:04:14it um but it's it's it's it's different
1:04:19so you can see it will now create the
1:04:21mappings it will take some time I'm not
1:04:23going to do it manually that the agent
1:04:24is going to create the mapping so once
1:04:26it's create the mapping I'll ask it to
1:04:28create the the identity resolution. Um
1:04:32again uh just for all of you right I
1:04:34will want to show really like what what
1:04:36would be my next step here or what is it
1:04:39that we're going to do here. So you can
1:04:40see the review you we have the reviews
1:04:42object, we have the contact stream and
1:04:44we have the app order object. Right? So
1:04:46we have all of these. So the next step
1:04:48is to actually use all of this and sort
1:04:51of map it to individual objects within
1:04:54um within Salesforce and sort of relate
1:04:57all of these um and then uh from there
1:05:01the idea is to create this uh identity
1:05:04resolution job. this one. Now I just
1:05:06need to wait uh for um the DMOS to be
1:05:10created here. So standard mapping
1:05:12succeeded. Now I need to map the app
1:05:14orders DLO. So now it's mapping. So you
1:05:17see the mappings is happening now. So
1:05:19once it maps my next step would be again
1:05:21as I said create the so you see the SSO
1:05:24individual is populated. Now let's get
1:05:26the app order dello and yeah next my
1:05:29next intent will be to create that
1:05:31identity resolution job and it's going
1:05:33to go go create it. Now I know a lot of
1:05:35you would have done identity resolution
1:05:37and you know the pain it's like you have
1:05:39to click at least 10 to 12 or maybe more
1:05:43clicks in here to actually do certain
1:05:46things in data 360. And I'm so glad that
1:05:49this MCP server is here and it's able to
1:05:52like you know sort of automate all of
1:05:54this and uh instead of me sitting
1:05:56through and clicking um the system can
1:06:00like reason through uh let's see if
1:06:02there are any any questions that are
1:06:04coming
1:06:06uh
1:06:08using headless 360 how to pass data
1:06:11effectively react as a front- end
1:06:13framework.
1:06:15uh very simple actually um this is
1:06:17something I've I've myself done it it's
1:06:20everything in data 360 can be queriable
1:06:23through Apex so Apex recently and let me
1:06:26show you guys this one um Apex connect
1:06:29query
1:06:31um so Apex um uh connect query for data
1:06:36360 Apex actually launched uh a new
1:06:41modern
1:06:42um query class actually that you can use
1:06:45it to just query. Yeah, it's called SFQL
1:06:49query name space. This is a new name
1:06:52space actually. Um I've used it often.
1:06:55So once you know your tables, you're
1:06:57just writing SQL statements within Apex
1:07:00basically very simple. Try it out and
1:07:02then you know you can build React apps
1:07:04or whatever on Salesforce and just use
1:07:06these to query u from data 360. We we
1:07:10made it much more easier now to just
1:07:12natively sort of use it. Uh let's see if
1:07:15there are any any other questions here.
1:07:17Um this is one. How is your diner data
1:07:21stream able to function when um fields
1:07:25have not been mapped? You might be in
1:07:27the process of configuring it. Um yeah,
1:07:30so the diner data stream. So first you
1:07:33create the data stream and of course
1:07:34it's not mapped but um you know yeah we
1:07:37need the mapping before we do the
1:07:39identity resolution that's why I'm
1:07:40waiting for it. Let's see if it is
1:07:42actually done. Uh let's see where the
1:07:45agent is.
1:07:47Um yeah agent is working through it. You
1:07:49can see
1:07:51uh
1:07:53most of them before creating. So it is
1:07:56so the other thing that Anushi saw this
1:07:59is one of the wonderful things is it
1:08:00actually goes back and looks what is
1:08:02there in the org.
1:08:04>> Yeah.
1:08:04>> Um
1:08:05>> that itself was previously taking me a
1:08:08lot of time right like whenever I need
1:08:09to go do context switch between what I'm
1:08:12doing and go back to the org and go back
1:08:14to my documentation it's saving me a lot
1:08:16of that time recently.
1:08:18>> Yeah and I think why it is doing is
1:08:20because I have this skill called reuse
1:08:22before reinvent. So this is one of the
1:08:25skills I wrought it in my harness and my
1:08:28plan is very simple is like anytime the
1:08:30agent builds something I wanted to go
1:08:33and see if there is anything it can
1:08:35reuse. It can backfire also sometimes
1:08:38like it will try to it will try to use
1:08:42um things that it shouldn't um but I
1:08:46have the skills. So so again uh this is
1:08:48very interesting thing right with the
1:08:49the skills concept here. Yeah,
1:08:52>> like yeah, it all depends on you like
1:08:54the hardness. I mean there is a thing
1:08:57like there will be a a baseline hardness
1:09:00that somebody can provide but then how
1:09:02you work um if you want to like like
1:09:07keep that non-sharable then you can do
1:09:10that as well like you can create your
1:09:11own skills and then use it within your
1:09:14harness. So that's that's very powerful
1:09:17um capability. So yeah, it's calling
1:09:20it's calling it's it's doing some some
1:09:22stuff here. Let's see.
1:09:25Yeah, it's it's doing uh it's basically
1:09:27going to create that object and then I
1:09:30think yeah the identity resolution I
1:09:32think it will it will create I I need to
1:09:35ask it to create identity res currently
1:09:37it is actually listing. So it's asking
1:09:40me see the standard DMOS so it kind of
1:09:43asking me the standard DMOS and the
1:09:44contact order mappings already exist.
1:09:47What should I build next? So create that
1:09:50DL. Yeah. Yeah. Skip further mapping and
1:09:54build the IR. Of course, we can't build
1:09:56the IR without we know that we have to
1:09:58map the the DLS onto the individual with
1:10:01the party and the rating. Uh this is the
1:10:04missing DMO. Yeah. So I'm going to go
1:10:06with that option. This is that loop.
1:10:09This is what I call it loop. If you are
1:10:11upfront like
1:10:14doing all the decisioning,
1:10:16you will never get in caught into these
1:10:18kind of loops, right? We can get rid of
1:10:20these kind of loops where it asks your
1:10:21feedback. But sometimes you can design
1:10:23loops where you know you are also part
1:10:26of the loop. So yeah, so now now it's
1:10:29going to build that uh DMO. I think the
1:10:32next step is that and if I leave my
1:10:34agent, it will next build the identity,
1:10:36resolution, segmentation, whatnot. In
1:10:38fact, you know, I had a spec that I I
1:10:42ran through this uh let me show you that
1:10:44spec. It's called Diner Winback and uh
1:10:48just because you know this whole thing
1:10:50can take like hours of long running
1:10:52task. We we are short of time here. But
1:10:56this is the spec I wanted to show that I
1:10:58actually ran in this arc. So I indicated
1:11:02all the source system. I exactly told
1:11:04what I need to do. Um I told it how to
1:11:07segment it. Um how to unify it. Uh and
1:11:11then my scope of the work and the test
1:11:12boundaries and my acceptance criterias
1:11:15and my assumptions and uh you know some
1:11:18things like you know I wanted to like
1:11:21keep a a companion doc of how to build
1:11:24like my own version of it. I also wanted
1:11:27it to inventory what is being built out
1:11:29so I can review it. Now this is another
1:11:31thing. If you are using a longunning
1:11:32agent, it will it can go and build like
1:11:35tens of components and then you will
1:11:37have to reason out when you are judging
1:11:39it. So it is always a better idea to
1:11:41inventory it um so that you can go and
1:11:44verify stuff because look verification
1:11:46is something that agent cannot do like
1:11:49it it can do in some way but it still
1:11:51needs like you know your your judgment
1:11:53here and that is very critical now.
1:11:56Uh I I was curious like do you want to
1:11:58take a minute to explain what is your
1:12:01process to create this spec and what's
1:12:03your process to evaluate the what the
1:12:05agent builds because sometimes we do
1:12:08feel when building is so easy it can
1:12:10feel slightly lazy to go and uh verify
1:12:13the work. I do feel that sometimes and
1:12:15and manual verification does not scale
1:12:18at the rate in which agents can build
1:12:20things right. So what's your process to
1:12:22evaluate whether the right thing has
1:12:24been done or not?
1:12:25>> Yeah, that's a good question. So here's
1:12:28the thing. I I I still think for
1:12:30enterprise applications, right? Um
1:12:33verification is definitely
1:12:37um needs to happen at human level
1:12:40especially for enterprise because
1:12:41there's lot of things at stake. Now this
1:12:43is also a judgment to begin with like if
1:12:46you know that this thing hasn't that it
1:12:49will not have that many impact then you
1:12:51can move really faster right like
1:12:53sometimes I'm not even reviewing the
1:12:55code I'm like yeah okay whatever the
1:12:56agent did let's just test it out like
1:12:58smoke test it's working fine but
1:13:01sometimes if it's very critical right it
1:13:05depends on the degree of criticality to
1:13:07answer in short right if it's very
1:13:08critical I I generally ask agent to like
1:13:12do some queries and you know automate
1:13:15that way like you know query it through
1:13:17this data lake objects and if I have a
1:13:20smoke test data given I will sort of
1:13:23manually do or at least manually keep an
1:13:27inventory that okay if this thing
1:13:29happens this is
1:13:31>> you know this is what I'm expecting
1:13:32right and then I ask agent to go and run
1:13:34that loop for me
1:13:36>> so it will go and run that loop right um
1:13:38so I can trust what the agent did is
1:13:41exactly correct or Not.
1:13:43>> Yeah.
1:13:43>> Uh but somebody like it's still somebody
1:13:47has to keep the system in their head
1:13:51like entire head because see it's like
1:13:53if you let agent sort of take over that
1:13:55thing from you. It will go and build lot
1:13:58of things that you have no idea and
1:14:00that's the worst thing like whatever
1:14:02agent build and if you don't have an
1:14:04idea about it that is the worst problem
1:14:06to have right. So that's that's the
1:14:09shift that is happening like because
1:14:11this implementation is so these agents
1:14:13can implement stuff very faster right
1:14:16>> um the the other part of the work
1:14:18suddenly becomes so much more important
1:14:20right
1:14:21>> yes
1:14:21>> um yeah so so that's the thing um that I
1:14:24would say
1:14:26>> and uh yeah I think the agent is still
1:14:28kind of working through it so look what
1:14:31it did um so here's the remaining new
1:14:34net mapping let me get the the field
1:14:36it's trying to All um for the run I
1:14:40expect a possibility category I'll
1:14:41execute it cannot map engagement you see
1:14:45it tried to map to the wrong thing and
1:14:48it suddenly realized it is not the right
1:14:50thing to do now remember like you know
1:14:52at this point that is the human signal
1:14:54is so much more important because for
1:14:56the next iteration I can put this in the
1:14:58skill in the skill file like whenever
1:15:01you are mapping these type of objects go
1:15:04map to this go map to that so now you
1:15:07know, instead of like your point, right?
1:15:09It will go back and sort of figure it
1:15:12out somehow.
1:15:14These agents are goal- driven. You have
1:15:16given it a goal. They're going to figure
1:15:18it out one way or the other. They're
1:15:20going to reason, keep reasoning, you
1:15:22know, like as you said, you know, it
1:15:23might have wasted ton of tokens here
1:15:25doing some unnecessary work. But it is
1:15:27coming to the same conclusion here. See,
1:15:29it was like contact point email mapping
1:15:31was created. It did create. It did
1:15:33create some mappings here. It's moving
1:15:35on but uh my point is same like see it
1:15:39created the mappings now you see it went
1:15:42and you know figured out that it needs
1:15:44to map to the contact point email. So we
1:15:47have a common email now right now it's
1:15:50going to create another object
1:15:52>> because it needs that review object and
1:15:54it's going to map the other fields to
1:15:56that and then once it's ready we'll go
1:15:58create the identity resolution rule. So
1:16:00as you said like you know sometimes the
1:16:02recipe is in your head and you want to
1:16:04just communicate to the agent. Sometimes
1:16:06you just let the agent lose, let it
1:16:08figure it out, you know. And both of
1:16:09these worlds have their own, you know,
1:16:12um, advantages and their own drawbacks
1:16:15as you said, like drawbacks and pros.
1:16:16And I think that is where the human
1:16:18judgment suddenly becomes like so
1:16:20important, right?
1:16:22Uh, yeah. So, it's going to work out.
1:16:24I'm going to leave my laptop. Um, you
1:16:26know, and, uh, as you can see, like
1:16:29these things can take forever, like, you
1:16:32know, uh, hour, couple of hours. One of
1:16:34the settings I did in my laptop is not
1:16:36to sleep. You know, this is the first
1:16:38change that happened after these agents
1:16:40and I saw how they work, right? So they
1:16:43can work and I think that is where the
1:16:44cloud loads are also going to become
1:16:46more critical like you can push this
1:16:48workload to the cloud and let all these
1:16:51things you know sort of happen in the
1:16:52cloud. I can see that u thing happening
1:16:55as well. U in fact not happening it's
1:16:57it's already happening. Uh with that
1:16:59actually I want to move back to the
1:17:01slides actually. I mean this is going to
1:17:03happen. Um I did run as I said with that
1:17:06spec and it did build end to end
1:17:08everything which was surprising. Um
1:17:11anyway so with that right uh I think we
1:17:13are running short of the time. I don't
1:17:15want to take much of your time as well
1:17:17Anush but for all of you right joining
1:17:19today a a big thank you right thank you
1:17:21for joining us and I hope you learned
1:17:23and saw this MCP. Now one of the other
1:17:26things I want is feedback. Definitely
1:17:28the feedback is there. Um, take a
1:17:31screenshot of this one. I'm going to
1:17:32keep it for a minute here. Um, this is
1:17:35all the resources that we have. I'll
1:17:38also try to do a trailblazer community
1:17:40post sharing all these things. Um, here
1:17:43we have blog post where we introduce
1:17:45this DX360 MCP server. Um, there is
1:17:49another code live sessions and um, uh,
1:17:53that is not the code live session. And I
1:17:55think this session this is on demand
1:17:57video that we have to learn about code
1:17:59extensions. I think they are also great.
1:18:01You can ask agent to write code to do
1:18:03certain aspects like rack pipelines or
1:18:06if you're building um transformations.
1:18:09Um and then um you know if you want to
1:18:12learn about cloud code, how to build um
1:18:15agent skills with cloud code, there is a
1:18:17link to there where we talk about that
1:18:20aspect of how to use cloud code to sort
1:18:22of build your own own skills. So all of
1:18:25these resources are there and then yeah
1:18:27of course um you know the other other
1:18:29aspect here I want you to remind
1:18:32uh the datablazer community right um
1:18:36please post all your questions Anushia
1:18:38is there as well
1:18:39>> yeah I'll be there to answer if you if
1:18:41you guys have any questions from this
1:18:43session you want some follow-ups you
1:18:45want resources that you have not gotten
1:18:47yet me and Moit as well as our
1:18:50engineering team will be in the data
1:18:52blazer community and we'll be able to
1:18:54help you out. You can start a post about
1:18:56today's session, ask any questions and
1:18:59we'll do our best to make you
1:19:00successful.
1:19:02>> Awesome. Awesome. Yeah, with that I
1:19:04think um you know we are end of this and
1:19:07again uh you know for all of you um I I
1:19:10hope this was educational and uh I hope
1:19:14you're going to play with the MCP.
1:19:16Please let us know if there is anything
1:19:18else we can help. Any parting thoughts
1:19:20here uh from you Anush? Uh no this has
1:19:22been great moit I think uh there's
1:19:24nothing better to use or learn about the
1:19:26MCP server than a hands-on session right
1:19:29this is not something you want to hear
1:19:31people just talking about it but
1:19:32actually showing how it works and I
1:19:35think you've done a great job in
1:19:37educating all of us in what your
1:19:39workflow was before how it's changed now
1:19:42and the way things are moving so I'm
1:19:44really glad thank you for everyone
1:19:45joining here and thank you for mo as
1:19:48well for conducting this session for us
1:19:50>> yeah absolutely Absolutely. Absolutely.
1:19:52And uh yeah um with with that right um
1:19:56again a lot of you are asking in the
1:19:58chat about how can we share some of
1:20:00these um skills and open-source work
1:20:03that as open source work. Um we'll
1:20:07definitely make all the possible
1:20:08attempts here to give you the
1:20:09documentation and the skills. Um and uh
1:20:13also the MCP server though is live you
1:20:15know make use of it connected to your
1:20:17coding agent if you're working in data
1:20:19360 space. your developer edition all
1:20:22orc also has all of these access. So if
1:20:24you're just learning please make use of
1:20:26it. Um and yeah um you know um we will
1:20:31make sure that we will give you all the
1:20:32all the resources. Uh and always please
1:20:36feel free to reach out to us and with
1:20:38that thank you. Thank you everyone. Bye.
1:20:43[music]

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 11,946 words across 1,747 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.