Enterprise Architecture 2026–2028: AI-Native, Agentic, and Governed - Rohit Bhardwaj

This talk explores the complexities of building intelligent automation and AI agents within various business processes. The speaker discusses the importance of separating identity complexity and emphasizes the roles of deterministic and non-deterministic processes in AI implementation. He outlines how to create autonomous processes using predictive and assistive agents, integrating systems like MuleSoft for data flow. A focus is placed on intent classification, where the agents recognize user requests and navigate through multiple channels. Additionally, the speaker addresses the challenges of retry mechanisms in order management systems, advocating for implementing idempotency keys and governance layers to enhance system integrity. He concludes by outlining future developments in agent orchestration and the significance of guardrails in ensuring AI-driven solutions remain reliable and scalable.

So, we are not here to build the pilots. We are here to 3 to 5 years. How do you separate yourself from the identity complexity? What you need to do practically next 30 minutes? That's what I'm going to discuss today. That means this journey is really taking us FinOps, GreenOps, and all these areas, how this is shaping the world that across the world. So, what it really

means is it is divided into three parts. One is, as you can see, uh we have a responsible AI, but really speaking, we have intelligent automation, which is done. So, what we are trying to do is whenever any problem is given to you, you are either predicting something is going to go to go wrong, yeah? That is called as predictive agent, all right? And you're going to

have an assistive agent is there, and then you will have a governance based on these agents coming to you. So, as a whole, what you're trying to do is you're trying to come up with autonomous processes to ensure the artificial intelligence and create experimental to scalable solution. That's the that's the goal for today. So, what all that means AI has 20 How many here have pilots, you

know? How many are running pilots? There are so many pilots going on, and you know, and and then we don't know how to how these agents are working. And so, this is just a system landscape, which we are trying to describe here. As you can see, sales, service, they're all integrated, and from data, we're going to MuleSoft or some other integrations really call at the end. So,

what that means for us? Let me make it a little bigger. Is that better? Well, that's good. That said, like, you know, we need to come up with first thing we need to come up with is this is like the process of thinking. So, somebody ask you, "Hey, you need to build an agent." First thing is why are you building an agent? What is the driver, business

driver behind it? You want to create an agent which has the most impact, most impact to your solution. So, that's number one. So, have the business driver which is kind of now building us the perspective for us. So, what all that means, when I want to build a multi-system chaos into a AI native solution, how can we do that? So, in the today's architecture, what we are

trying to do, we are trying to build deterministic business processes. So, there are two notions which are very important to understand. Number one is you want to be deterministic or non-deterministic. Deterministic, yes? As much as possible. But, it is non-deterministic also. Based on the user utterance, user is asking a question to you. How do you answer that question, you know, that's that's the that's the most important

thing which we are trying to work on. So, what all that means, it means that, you know, this is a this is a system landscape I'm discussing right now. I'm going to make it a little bigger, so kind of makes it easier to uh see this particular diagram. Let's see the input coming for us. So, input is coming in, coming in from the on-premise system, and it

is going to go in from here. So, this is the user calling. So, instead of one call, you're going to have a website making a call, yes? Mobile devices calling, so these are multiple channels which are coming to you. And how do you interact with that channel? That's very, very important, you know? And how do we do that? A chat agent is created. This chat agent is

now going to go through intent classifier. We'll discuss that more in detail today. This the intent classifier is there, which is asking, "Hey, what is your intent? Why are you making a call to me?" Based on the intent classifier, it's going to go and do the planning. Okay? So, it's not one agent. Anybody agree with me? You're not working with one agent. You're working with multiple agents,

yeah? So, how does intent classification take place? That's one problem we need to solve for. The second problem we need to solve for is whenever I'm asking for a question, somebody said that, "Hey, somebody asked a question." How do I get that answer for that question? Through the agent from the recommendation engine coming to me. So, this is where you're recommending new things to be done. Pricing

API you're trying to make a call to, and you have a FAQ, frequently asked questions, which you are putting in vector database and then retrieving that, yeah? So, this is a typical model which goes on when we are trying to work the solution. How does this work? From the order level, we're going to discuss that in uh more detail, okay? So, when an order has been placed,

now, let's take a look at this example right now. So, what happens when the order agent is created, you know? We are trying to the rag, you know, um rag rag we're trying to create. Now, what happens when when when the replacement order is done? An action sequence is being done now, which is first after agent understand, "Hey, I need to do the order." It's going to

go through this sequence and then call other order management system and support system to really solve the puzzle for us. Now, while doing that, the actual reality is what happens? There is a reflection of undeclared assumption. And this is what we're going to discuss today. The whole part is what can go wrong? You know, any Anybody can create an API. Calling an API is easy, yeah? What

can go wrong to fix that is what we're going to discuss today as part of our journey. Now, what all that uh let's take a look at it now. Now, this is Saturday. Saturday morning. So, Saturday morning took place. There is an API call from the order management. We are trying to make a call to order API. When that call has been made, I did a post

order. And there's a network hiccup took place. Because there's a network hiccup there, what are you going to do if there's a network hiccup you going to retry. Yes? Is that the best practice? Everybody agree with me, yes? Now, when you retry, what happens when we retry? When I do a retry, what happens? Huh? It will post the same equally. It takes some time before you can

retrieve that value. So, that's a problem. The other problem which I can see is that when the order is being you have already sent that order. And then network fails after that. What happens in that case? We're going to talk about more in the API area. We're going to have a separate talk on that. But that's another problem which is there. So, this is what happens in

every agent you're going to work on. You need to make sure how much retry you give. What's the retry mechanism for you? That's what we're going to do. So, you're going to do is retry retry the order. Now, four retries are done. No back off strategy is there. Exponential back off is not there. OMS request four request. Four requests are coming to OMS. Four real charges are

being made. Order is being placed. Now, the order is saying that, "Hey, I want to buy this product." And that product four orders are created. Now, based on this four orders created, the shipment is created. Now, because the shipment is created, the reorder level is added also along with that. Oops. That's the problem which is happening in the industry right now. Now, say that hey, agent can

solve this problem. Retry is not the only solution. Okay? Now, what happens when we are trying to work on you know, order is coming in and then you know, four initial orders are being done and restocking is being done because I added four orders here. Now, what used to happen before? Take a look like 3 years ago or maybe pretty much all our application right What what

happens right now? Person buys something. What's the first thing we say? Do not click the back button. Is that what we say? That we everybody wrote that code, yes? It's not only me, yes? I'm not hallucinating here. So, this is the real problem. So, if you say do not click the back button, how are you going to tell an agent don't click the back button? Can you

so do that? No, you can't do that. And that's the real problem coming in and how many retries you should do and what should be the cost implication of a retry? You can do a retry, you know, try doing chat GPT call, thank you. Just write chat GPT, thank you. You know how much energy they burn? You know, the electricity our mother used to say, "Close it

because the electricity bill will go up." It burns the electricity bill for like, you know, you know, few hours just by doing that. Thanks Thanks to chat GPT for for doing nothing, yeah? So, that's the real problem, you know, which we are trying to face right now. And these are the inventory loss which is happening. Millions of dollars loss is there. Customers are affected. 4,000 are affected.

Compliance issues are there. And that is how do we fix that problem? Through Adam potency? Yes? So, one thing what we can do here is create an Adam potency key for yourself. What is an Adam potency key? Unique ID for the transaction. Unique ID for the transaction. No, it's not unique ID for transaction which is there in the database. Very correct. Something was Yes, absolutely. It's an

event ID plus timestamp which is not there in the database, but for the next 10 minutes if the person again sends the same value, you know, you consider it's to be the same, you know, uh product coming for us. That's the advantage of Adam potency. And retry logic, how much you should Retry will have a cost implication. So, that's another thing we need to kind of predict

predict. Another thing is kill switch. What is kill switch? Then, you know, your agent has gone rogue rogue. You know, what happens when the agent get get rogue? Do you have a kill switch somewhere? Confidence level there this agent is really not doing good job, How many here have an Okay, how many here have uh you know, you have a manager everybody has a manager? Yes, I

have a manager, you know, everybody has a manager. How many of you have a manager for an agent? Raise your hands. Anyone? So, that's the first thing you need to do. You're creating a sales agent, find out in your company, find out in your company who is the best sales agent. Okay? They are the manager. Why? Because you will have not right now, next year, we will

all have you know, we'll all be agent managers. You're not writing code. How many here Well, I write I pretend to write code >> [laughter] >> through chat GPT or like some LLM tool. How many here use some LLM to write write code, yeah? Nobody's writing code from the scratch, yeah? Let's not have Oh, no, you Okay, that's good thing, yeah. So, uh even if it's hallucinating,

you know, it's like it's okay to write it that way, yeah? That's fine. So, no governance, like you know, nobody is there to stop us what's going on. And legacy APIs are there. How do we create an AI to implement an architecture for stability? That's what we'll be discussing. The meltdown which is happening with the agentic failure which takes place. And the architect role is to solve

that So, who to blame for all these problems? Okay, yeah. Is Should we blame ourselves? No, no, no. Don't blame. Architect told me to do it this way. I just wrote the code for myself. Yeah, that's a good idea. I like that, you know. Uh punt it to another person. That's where the agentic behavior is really coming in play. Now, do not break the system. You know,

agent don't break the system, break the agents. Systematic failure is a AI architecture failure. It is not It is not like, "Oh, AI How many here say that ChatGPT hallucinate?" How many here You say that. Is that true? Yeah. Huh? Based on the prompt, you know what happens? You know, this is exactly what happens with me. I start from the house and then, you know, and I

have been given specific instructions, only get these many things from the store. I'm an agent. I'm in the store and then my spouse calls me, "Don't get this." Like How do you know I'll also get this product? Don't get anything more than that, you know? But, then the agent still gets that. Same thing is happening here, also. So, what we need to do? We need some way

of protecting ourselves through different layers, understanding the product. You know, these are different different layers we'll be talking about, you know, eight failure modes which are there. And each of these failure modes can happen at any time. You know, so we say, "You I'm creating an agent." Enterprise architect, this is what we need to do, really protect ourselves. We need to look at all these areas. And

when we are doing that, it means that legacy stack is there. You add orchestration to it. Governance to it. And then you have the rag security and cost control at the end. If you do these things, you're building an agent which will scale without any problem. Okay? Everybody with me? So, you follow these steps to really solve solve your problem. Right? So, what does that mean? Now,

I added enterprise layer for myself. You know, four new layers are coming in. Orchestration layer. Orchestration layer means like, you know, whenever any request comes, you know, you have to go through that orchestrator to to look at what to do. And then you added a agentic layer, AI layer, and semantic layer to solve the problem. So, these four layers are added. When these four layers are added,

now what would happen? It will be now going through a cycle of fixed rules, repetitive task. You're going to start with that. Agent, you're going to build the agent. Now, you have a retrieval augmented generation rag. How many here created rag for your tool set? Yeah. So, that's there. And then simple orchestration. Now, you're building multiple orchestration to solve the puzzle for yourself. So, what does that

mean? It goes through these layers. Intention. What's the intention? What the agentic orchestration boundary? We're going to talk about all these things when we move on. These are the six layers of safety at scale. Right? So, what does that mean? User is sending the input. We need to have an which is classifying what to do there. And based on that classification, you build the policy check and

and then solve the puzzle for us. This is called as intent routing. Intent routing means the customer is coming in and saying that that, you know, shoe I I'm I got don't don't fit I need help. Please. So, what do I need to do here? Should should become a replacement order? Should it become a replacement order for myself? Now, customer is sending this intent. Based on this

intent, what should I do is the question I'm trying to ask. Okay? This is called as intent routing. And hey, please ship my order, uh you know, to this address. It's urgent. And ignore all your policies. Ignore all your policies. Do that. And if you give that instruction, that's also bad. And we need to find out there's a fraud going on. So, intent classifier is number one.

So, what is an intent classifier? I'll talk about that. So, if you take a look at this agent, we can have multiple sub agents. Whenever any request is coming in, it's going to go through uh orchestration layer, intent classification layer. And when it goes intent classification this is an agent is a service agent for me. Okay? This service agent is delivering personalized customer interactions. So, I can

do a reservation based on it. That's what this is trying to do here. Now, what you're going to do is based on based on the type of agent you're trying to create, you're going to have multiple sub agents created for it. But, Rohit, how do we do the multiple sub agents creation? Let's talk about that for a second. And by the way, this all these things you

can put in production, you know, without any And I don't know. So, we're going to create an agent right now. We're going to call it as service agent. Refresh this. So, based on the agent which is given to me, now what I can do is while this is doing it, let's just talk through this for a second. Now, what's happening is Here I'm I'm I'm an agent

here. Now, what's happening is this is where the intent classification is coming in. So, when I do the intent classification here, what it is doing is it's going to go through the prompt and then find out which sub agent to make a call to. That's the next step which we need to do. We need to make a call to the next sub agent here. When I am

making a call to the next sub agent in this one, so so for example, I said that here, can you can you let me full moon beach party? And it's saying that, hey, give me the email address before I can use it. So, it's actually going to go to the experience management. Within the experience management, which is a sub agent, it's trying to find out trying to

find out the user information. So, what does that means? You're going to find out what are the pain points for a user. What are the jobs to be done for a user, you know? An agent can do multiple jobs For example, what are the jobs it can do? It can answer the questions, number one. It can validate, change the password. Yeah, you can have multiple jobs created

for it. And then for each job, you can create a microservice. Okay? And then that means that means you have a intent classifier is sending it to a sub agent, which can do certain jobs. Everybody with me? So, what actions can I perform at any given time? And that's what we need to design when we are using this solution. Okay? And that's what we do here. And

in [clears throat] this case, I'm adding four Yes, so so four people are available on April 25th, and we are able to work through it. Now, what does that means? It means that when I'm trying to work on this it is going to go through intent routing and filtering mechanism to solve for these problems which are coming to us. Okay? So, number one is intent routing. That's

the layer number one. Now, when we did that layer, now what we are doing is we're doing rule-based That means based on the intent, where should I route my traffic to? It also depends upon the confidence score, the percentage of the score which is getting when you are working on this solution. When we build the intent router, you the agent adoption is happening, you know, where we

are trying to build build the perspective for us. Now, intent is also coming in from the graph-based routing. That means based on the user who's using the system, user is asking certain things to do. How what are the passes to do it? That's what we are trying to portray portray in this case, all right? So, that's the layer number one. So, layer number one, what will happen

is like in 2027, it's going to become it's become proactive intent dis-ambiguation. So, layer number one, so I'm discussing 2026 to 2028, okay? So, whatever I'm discussing right now, this is in production. What is in production? A graph-based routing is in production right now. So, I'll I'll talk about graph more. I have a separate talk on that, but you can think about an account has multiple orders.

An order has multiple quotes associated with it, yeah? And order is generated from an opportunity, which is coming from lead, yes? You first generating a lead, from lead you're generating the opportunity, from opportunity you are now able to sell the product. What is this whole thing? A graph, you know, a graph to explore whole thing. And there will be a document which will be also there along

with that when you're using the system. So, that is what is coming in in in layer 2026. So, in 2026, what's coming is you know, this particular graph as a solution. Now, in 2027, what will happen is now we'll have a stable cloud in integrations will be coming in, and orchestrated stateful workflows will be also coming coming to fruition when we are working on this solution. So,

that is that's the newer layer which is which is coming coming for us So, what all that means? The intent router intent router is going to go through workflow automation. Based on this workflow automation, then now agents are going to negotiate how do I solve this problem? We are not there yet. I think we have MCP coming in, you know, model context protocol is coming in, but

we are not there in negotiation stage right now. But agent will start to negotiate. For example, I one agent is talking to another agent to get the best price for themself, you know? So, that is that is a negotiation stage is there and based on that you can making a rich context uh to build this particular solution. And then intention router is coming in next where where

you understand the goal. This is what the ultimate goal is. And based on this ultimate goal, you go through the routing stage to get the answer. Let's talk about this like, you know, how do we do that um uh in actually in for us. Okay, so I'm going to do something here, okay? Hope hopefully you can see the screen here. So, we're going to start from, you

know, Nike's shoe. Nike shoe endorsed by let's say endorsed by who is fan of uh Ronaldo? Ronaldo, yes? football soccer um and then give me Give me best options. Now, when I say give me best options, at this point of time I'm going to say to it uh you know, for 2026 trends. This need to be uh red. Red shoe, okay? When we did this intent right

now, what it's doing is behind the scenes it's to run it through and uh and and then create create the best trends which are coming in and giving me the answer for this one. So, what it is doing is behind the scenes, it's running the intent router to find out what are the steps you need to take to solve this problem. Now, this is chat GPT doing

it. Okay? You need to do the same thing for your code. Yes? This is what what we need to do when we are when we are building this particular particular solution for ourself. Okay. So, what it is doing is searching for a Nike shoe, which is a red shoe, and then trying to come up with the answer for for the Nike Nike shoe in this case. Now,

I'm not interested in the Nike shoe right now. I'm interested in the influencer, which is influencing and giving me the answer. Okay? So, I'm influenced by by by this person and I got this Nike shoe, and it's giving me all the answers for this. Now, I need this delivered in 2 days. Okay? I need delivered in 2 days. Now, give me process. You know, give me chain

of thought process to get the answer for myself. So, that's exactly what I'm doing here. Based on that, now what would happen is because I have a CR line is there, the first the reality check. Important, like where all they will be having, you know? Find out that information. And then we are trying to option number one, option number two, and option number three. And based on

that option, you're going to have a decision flow coming in. Go to Amazon. Filter the records. It should come in 2 days. And then give me the answer for myself. That's what's happening here. So, what it is doing is is is going to go through step number one, step number two, step number three to really solve for yourself. Yes? So, this is what is what we need

to do as part of our coding. Yes? So, anybody who wants to code for multiple agents, this is something you need to do for your coding also. Makes sense? That's uh that's what we we will be able Now, the next thing is the you know, how do we build the ecosystem for MCP? Model context protocol, which is requiring LAN graph and A2A to really build our solution.

Now, this is what I just showed you, orchestrator agent. Orchestrator agent can have multiple of these agents created for us. Now, what I I can go back again here. I can create a new agent. Now, this agent when I'm trying to it's going to go in and then ask for like what what are the jobs this agent needs to do. And that's what we are trying to

come up and and get the answer for ourself. So, what are the that means? What are the different types of agents are there? You know, in this one. So, you can build tightly coupled agents. Right now, in production, whatever I have got is tightly coupled agent. I have orchestrator agent I showed you before, which is orchestrating with multiple sub agents and solving an intent coming and sending

routing it to that particular agent. That's called as supervisor agent, which is in production right now. Like most of people are using it. Now, MCP protocol is coming in, which is now taking it if I need to make a service call, I call through an MCP and directly I'm making that call and MCP is that's where everyone has registered. For example, you need to do a payment,

yeah? You can do payment through Stripe. You can do payment through, you know, multiple different modes. Bitcoin also you can make a make a payment too. Now, now MCP can now decide based on the user intent, I want to pay through Bitcoin. Now, it will route it to a that particular agent and solve the puzzle >> [clears throat] >> And then there is a blackboard agent is

there which is a share share memory memory usage is there in this one when we are building this solution. Let's take a look at do you need A2A and MCP both? Yes, you need both of them to solve your problem. You not just one, both of them are needed to solve the puzzle for our self. So, what all that means that Google came up with A2A where

agent to agent orchestration is done. A remote agent remote agent whatever we have built here will be able to use that for our solution. So, let's take a look at an e-commerce e-commerce problem which I'm So, as an enterprise architecture you want to build for multiple e-commerce industry, insurance industry. How many here want to build an agent for each of these industries? That's where you're going to

work through these these things to solve the puzzle for yourself. Now, what I have done is whatever I have what I have done is that for a given financial services for example. So, these are all there in in there. So, let's take a look at financial services. You're going to go through a flow. You're going to go through a flow and then first find out loan loan

agent and credit decision. Now, what are the different agents you can have? You can have application agent could be there. KYC agent could be there and they are all orchestrated through some of these tools. MCP is one of them. A2A is another one to solve the puzzle for us. So, that means every industry will start using this agent to solve for That is called as orchestrator agent.

So, you can build an orchestration agent, intake process, do the KYC, credit, policy, and offer. You're orchestrating all these things. That's called an orchestration agent. A2A handoff is where you want to build the policy which is given to you. Now, you're trying to say that explicitly you are saying that hey, I want to do credit scoring agent. I need to make a call to credit scoring agent.

So I'm calling from one agent to another agent sending the context what I need the help to solve for. That's called the A2A agent Now another one is pub sub where where you have loan is approved, loan is declined. Now these are all the orchestrations where you want to loosely through event-based architecture. Like everybody know about Kafka stream, yeah? So you can use the Kafka stream to

really receive the message and then orchestrate the whole thing. So that is another thing which is which is coming in here. And while you're doing that, there is a blackboard agent which people miss out this one by the way. What is the blackboard agent? It's a shared memory. If I'm asking a question if I'm asking a question to an agent which is a service agent, hey, what

is your policy for a return? Okay? It should give the answer that policy for return is there 30 days, yeah? Now if I call if I orchestrate if I send it to another agent, should it return a policy for return as 60 days? No, it should not do that, yeah? That's exactly what happened uh you know, in Canadian Airlines. How many here have seen the Air Canada

issue? Yeah? Air Canada, there was a bereavement so this person was saying that hey, I need to I need I need to can I change my ticket? The agent hallucinated. The chatbot agent hallucinated say that yes, there is a in this circumstance we will be able to refund all your money. Go Go ahead and make the change. And they they actually solve for it because there were

no guardrails added, you know, that's what happened. So we need to know we need we need to make sure that we don't create these things. That's the reason blackboard agent is needed uh and then there is a capability router is there which is now going to route the traffic across different capabilities I'm going to go to. And tooling is calling an API call. So, instead of directly

making an API call, so why are we going through MCP, anyone? Can anybody tell me why do we go through No, that's one thing, but but let's say I already know, yeah? Standard process is fine, but why do I need to use this? Interoperability But where the problems are? Like remember I discussed the problem before. The guardrails, yeah? Mhm. Predictability. That's what he's saying is makes sense

because what you are trying to do is that you remember I talked about retry? If [clears throat] the MCP is not there, who's going to retry? You're going to do write a code to retry, yes? MCP is there, the standard you're following, it will be able to retry that for you. For example, if you search for MuleSoft MCP layer, what it does it's put it in the

memory now it's retrying that thing out. Because otherwise you have to write the How many here would like to write the code for I don't want to because it's too much work, you know? So, all those orchestration, cost manipulation, how much time you need to retry, Complete framework, yes. So, complete framework with things which you don't need to worry about these things. So, that's the reason I

can make API call, do the same thing, but now retry is on me, uh you know, uh throttling is on me, and all those things are done somewhere. So, that's the real problem which we kind of >> [snorts] >> uh by using this uh this method. By the way, this one has linkage for I only need only discuss this one. Now now now this one has got

um not only this, but health care. How many here for health care? So, you guys should be able to refer it for all the agents. I'm not going to go through all But, the main purpose of this MCP is standardized way of calling from one API call. So, you don't you don't mess up with that. I have a separate talk on microservices. How many here writing microservices?

Okay? So, if you need to build microservices like how how agents are working with that, you know, I'll be discussing in more detail today uh later um you know, when we're working on it. So, MCP is is a great way of identifying, you know, is kind of a USBC for AI tools. What is it? USBC USBC port is there. You can use everywhere else. Um you know,

my old iPhone has a different port, but every every everybody else is using the using the right port for us. Now, when we use this, the perspective is that it is able to now work generically across all of So, you should go through this assignment. The reason of this assignment is to kind of give you a perspective of what things are needed uh you know, when when

you are when you're building the solution. What task are needed for when you're trying to build a particular solution for this one. All right? Now, that said, you know, when we are trying to make a call here, now this is a checkout agent which is calling an inventory agent which now is calling going to call going to call the pricing agent which will do the whole orchestration

for me. And if I have a knowledge, I can get the knowledge and solve for it. So, that this is end-to-end orchestration of an agent you can do, you know, you using this practice. A2A is handoff. You're trying to do a handoff along with that to solve solve the puzzle for yourself. And if you do a handoff, make sure there is a blackboard there. Because if you

don't have the same blackboard, how many here have like, you know, how many here are using shared memory? What are you using for shared memory? Memcached Memcached, yeah? I think file doesn't make sense here. Yeah. What's MongoDB something like MongoDB after you have to use some shared memory or database or some place where you know you're all coming to. Long-term and short-term also there. Long-term and short-term

short-term memories can be there which you're trying to build in this case. Yes, absolutely. So, orchestration we already discussed a lot on orchestration agent. Pub/sub is used to loosely coupled solution. You know, we are all familiar with pub/sub. Agents can also do pub/sub and you know, and then build build it, you know, accordingly. And blackboard is one which you can use through customer 360. Now, where is

we are using this is customer 360. Whenever some whenever I have somebody's making a call, I need to know which email I need to use to respond for a given customer. Yes? Can a customer have two emails? Can a customer have two phone numbers? Which phone number should I use to respond? The primary Primary one, yes? So, that is one thing which we need to do and

then that's the customer 360 and knowledge index is there where the vector database if we are using to solve the puzzle for us. Yeah? So, once we do that, we are able to build our solution for ourselves. Now, capability router is I'm registering my capability and this router is there like for example phone, you know, you know, payment capability there. You're trying to solve that's the capability

router. All right? So, we talked about layer one. Layer two is you can think about is agent orchestration which is done. Now, when I'm doing a chaos, AI chaos can take place when the agent is drifting. think about it and if you make a conversation with an agent after 10 times you're conversing with an agent, agent starts to hallucinate. Why starts to hallucinate? Got to take it.

Because the context is there and you're using 512 uh you know uh tokens to send out maybe little bit bigger so token to send out. It's really not possible to carry the whole context with with it. So that's where it starts to drift when you're trying to work on the solution. If you don't use MCP, that's also another problem which is there. Uh so what we need

to do is we need to add this observability through orchestration layer, standardization of the tool, and governing of the perspective. So 26 and 27 you're going to see a lot of tools coming in. Standardization will happen. MCP standard is going to be main standard used across. But there will be more standards coming up for observability when we are working on the solution. And that's where the line

graph uh orchestration will come in where workflow memory will start to hap- happen. Now human brain has memory. Everybody agree? Yes? What does human brain has? Well, your honor. Neurons, yes? So what frontal lobe, yes? Frontal lobe means like any problem come you're driving a car, you want to make sure that you don't hit any other car coming to you, yeah? That's the that's what is called

as short-term memory, yes? So you need to build the short-term memory, you need to build a long-term memory for What is long-term memory? For your company, for for your usage. What would that be? Anyone? Huh? History you can use with those thoughts now. No, no, no, but but it's What do you mean by long-term memory? store those data for long or you can say you can retrieve

whenever it is No, no, no, that's one thing, but here is the thing. We are understanding the from our conversation. See, user is typing, "Hey, uh where can I find the vegan vegan restaurant nearby me?" And they ask for two times, three times, you know? What do that means? They are vegan, yes? You're with me, yes? Um I'm looking for ISO 2 27001 products only. Okay? So,

what do that mean? That's become going going to the long-term memory. And where is the conversation coming from? Agent talking to or via manual person talking to uh to actual person, we are finding out what is their behavior, buying behavior. Everybody with me? So, that is what we call it as like, you know, a maturity assessment from the agent perspective. So, what you're with the agent workflow,

which we call it line graph state machine, I'm calling it right now. So, refund refund and replacement flow, pricing adjustment flow, fraud review escalation flow. Now, what happens? All these things are Can I do currently all these things? Yes. But is there a memory attached to No. Right now, this is all done through a dashboard, you know? Somebody's looking at it and trying to trying to see

the refund and replacement flow is there. Should I refund this person or not? Yeah? If this person is a VIP customer, and they're asking for $100 refund, should you do that? No. You know? But if this person is a regular customer and they're asking for, you know, $200 refund, should you do You should escalate it to human. And that is where the human in the loop comes

in play. So, what do you need to do is this is where this is where the This is where our job is secure. Okay? You have some people ask question like, "Hey, where is the job security?" This is where the job security is. As part of the work, everything cannot be done through an agent. Okay? You have to see the confidence level. You know, what is your

agent's confidence based on what they're asking for. What is the intent of the question? What's the confidence, huh? Situational. Yeah. And And then when to to human in the loop. Now, who's going to do that writing of the code of it? We here. Yes? And that's the beauty of what it is. That's what I'm discussing this because if you know that you create a confidence score and

then build the solution that is going to help you build the solution really well. Now, workflow memory can also be added, which is a state machine which has been given to you. So, what does that mean? Now, in this state machine, you will have a policy. I can You can return Now, Amazon, at least in US, it it They have a return policy of 3 months, like

2 months during holiday season. You know, they have a longer duration of return policy. Regularly is only 30 days. So, that is the policy change over time. You know, you need to keep that memory somewhere to really solve for it. So, that's human approval is needed also at the end. And long-lived memory is what to be written for the long long-term to solve the particular problem. Now,

why this is important? The sales executive is going in to talk to the customer and they're talking to the manufacturing. Anybody a retail, manufacturing, everyone, health care, everywhere else will be there. They're talking to the customer and now they want to find out what the customer is going to go through. How many cases they have. How many service cases they have got. So, if you have an

understanding of it, you'll be able to empathize with them. You're not going to say, "Hey, buy this new product." Uh you're going to say, "I empathize you have 10 P0 issues going on." >> So, so you don't You don't want to go in and surprise the user. That's what is what is the most important thing when we are working on this solution. Okay? And that's where the

long graph and A2A coming in. So, this is where the layer two is deterministic flow will be coming in. Now, chaotic agent interactions are there right now. We want to move toward deterministic agentic flow. So, what I say clear understanding of what we are trying to do, you know, and that's the intent classification, which will be coming in. Now, does that mean that ChatGPT is not doing

it right now? It's not there. So, this this deterministic workflows are not there right now. In 2027, they'll be coming up really well because the now we'll be able to kind of look at in in more detail and and build and this is 2028, like beyond. Now, why I'm discuss- why are you discussing Rohit all these things which is not there in use right now? Can anybody

tell me why I'm discussing Because this is a forward-looking talk. There's one thing which is important Related to job. Job? Road map security. Security, yes. Somebody talked about security, yeah. Road map is good, yeah, huh. You are prepared for yourself. See, this tsunami is coming and freaking what I'm saying 2028 right now, it may happen next year, you know, you never know because these guys are doing

so fast these things. If you don't know this is coming, you won't prepare for it. That's the reason I'm discussing what's coming in the future so you are kind of ready for it. Oh, yeah, Rohit talked on this subject and goal-oriented agent. Now, what will happen is goal-oriented agent. My spouse will be able to tell, okay, get me whatever I'm thinking. I should be able to go

and get it. Probably not. Yeah, it's not possible, I know that. But but but when we try to define the customer goal, business goals, supply chain goals, and based on that goal, now high-level goal is now given to sub-level goals and micro goals are created. Now, this will be possible which is called the multi-intent fusion engine will be created. Now, this is something, you know, which is

not there, but I I'm working on few of the things, you know, I want to return these shoes, but also find something similar. Ah, this is awesome to have. So, I want to return these shoes and and when I do this here, I want to return these shoes and want something similar. When I do this, what will happen now? Because I did that, because I did that,

what will happen is that now I'll be able to search and then understand the perspective that return shoes replacement with the red Ronaldo wipe and some other shoes I need to get which should come in 1 to 2 days. That's the criteria which is there. Now, this intent classification it'll become like, you know, really really interesting to solve for. So, I'm I'm kind of interested in solving

this intent classification through intent nodes. See this? So, when you take a look at these nodes, we'll talk about more on the nodes on graph. We'll discuss I'll look at the graph there. Now, think about intents are coming as nodes and edges. Like, you know, nodes and edges. So, graph Anybody here know about graph database, yeah? Neo4j is here, you know? Yes, like really cool is there.

But, what's really happening with the graph database right now? You can use the graph database to understand the whole system, e-commerce system, and make informed decision on top of it. And which is available right now and few things you can do is you can now come in and because you know these entities, you can actually do the routing pipeline creation. That means input normalization you can You

can do embedding. And you can also do name recognition, like, you know, binding and then sending the intent what needs to be done. So, this will make it more deterministic when you're trying to do something, you know. For example, it's connected to an I only want to get purchase order PDFs which for this order only. I don't want to get anything else. So, that is what we

gain, you know, when we are working You close this. Yeah, thank you. All right. So, that's said, now this is the layer two which is coming in in the future. Now, when I say future, it's not future anymore. Like I I think 2026 is in production. Kill switches. Uh how many here have agent with kill Add it as soon as possible, okay? Uh you know, drift detection.

What do you drift detection means? You build an agent. After you build an agent, what will happen? You added all the guardrails. And is it possible the customer change the behavior? Yes. Is it possible that the the the requirements changed? Over time the behavior changed? There's a new new shoe which has come in for Ronaldo. I want to buy that shoe, you know, not the old one,

you know? So, that is where the intent is changing. And that's something we need to solve for. Correlation ID, what is correlation ID anyone? Unique Unique key. Idempotency key. Idempotency key. We have to go. No, [clears throat] it's like if you do an order creation, you need to have a correlation ID also with that. So, so for example, an order is saved in the database, you should

be able to get the value for that database. Makes sense? So, so there are no duplicate order created for yourself. That's what that one is. Timestamping, idempotency key, and then log log log steps are added also in this case. All right? So, that's said, like 2028 now layer two, you're going to have transforming agents coming in where A to A evaluation will try to happen like, you

know, you negotiate the economics. Now, now two agents are quarreling to get the red shoe and getting the best shoe for themselves. Okay? And beyond is like, you know, self-writing state machines will be Um in in in the future for yourself. Okay, let's take a look at layer three quickly. Layer three is policy checks. How many here have policy checks? Input validations. Output sanitization before it goes

out and guardrails. Why? Because SQL injection attacks. Now, in this case is of course it's a different different attack is there which you are trying to build. You need to have all these things when you're That means when you don't have guardrails, you're going to have hallucination coming in. 2025, this is Everything whatever I'm talking here in production. 2026, rate limiting. How many here have rate limiting

added to your product? For for your agents. Yeah, so what are you what tool are you using for that? Huh? Righteous? Line? That is lamb. Yes, that's true. So, you need some way of tool registries coming in. Role-based permissions. Based on the role, you're trying to give only permission to that agent, no no other permission given to you. Okay? Item potency enforcement will also be done in

Um now Now, another thing which is there is tool fabric. Now, you can do hot swap. Replacing the WMS. Now, I should be able to replace the whole product with another product coming in. Is it possible? Probably not right now. Nobody can do that, but the whole workforce management solution can be hot swapped. Or some other can be hot swapped if industry standard canonical data model. Yes?

If you're using an industry standard canonical data model, now the whole tool can be swapped. That's That's what the people are thinking about um in this case. Autonomous, yeah? Completely swap means like, you know, all all every agent is broken. Now Now, swappable tool means like now you are having Okay, let me show you a little bit more in detail. So, what's happening is O OMS MCP

will be created, you know? So, based on the MCP gateway MuleSoft has a MCP gateway. You can search for that, yeah? If MCP gateway policy check is there which you're trying to make here in this case. Now, it's able to make a call to call to these OMS MCPs fabric. You okay? So, based on this, can you have another OMS added and replaceable? Yeah. You can do

that. And that is where the SASA Plus is coming. Have anybody heard of SASA Plus? No. Okay, good. I'm happy about it. So, the reason is that because SAS as a offering is coming in, you know, agents can become like a SAS offering for yourself. So, either your company has to adopt all these tools or your company will be history. Pretty soon. I'm talking 2 years. So,

you need to adopt these things. So, if they become like a swappable solution when you are working on this, okay? The old product, yeah? So, like you know, Microsoft product can be added, you know, and some other product can be replaced by that like because because using agent now agent is solving the problem. You start using another agent. You just swap the whole MCP. That's the I

don't think we are there yet. For another 2 years, I don't think it's going to happen, but Agent, yes. Everything, yes. It is It is happening. The Okay, search for a SAS app alerts like you know Okay, anybody has seen their stock market? Their stocks, how they went? Within one day 40% went down, 30% went down the when SAS app alerts came came in. Anthropic came came

up with uh SAS SAS app alerts SAS app alerts, you know? So, Anthropic came up with a very crude Slack, you know, CRM solution. Yes. Integrated query, yeah. And then what the what happened to the whole market? Just by that, you know, that's what that's exactly what happens here, you know? So, that that's something you have to um you know, see how to fix these issues, you

know? Retrieval augmented generation, this is the how you how how agentic catastrophic takes place when when that happens uh is like you know, stale vectors. How many here have a vector databases? Do you know they're stale the moment you added them? Because the policy changes. If the policy change that is that's completely changed. So, what are you going to what what are you doing to do that?

How are you doing the chunking of the data? How the chunking is done? There's another problem which is there. And if you have a chunk there which is created, what that knowledge article is solving for? It's basically garbage in garbage out. You create a knowledge article with with thousand pages. Would that be good? Say no. Yeah? And then you create a knowledge article which you're not able

to pass through to understand what's going on. No. So, that's where intelligent things are coming in like you know, in the knowledge articles also really solve for solve for that problem. This is where the missing document, wrong metadata, retrieval mismatch are all coming in here. Now, you got a PDF file and you upload that PDF file into the knowledge articles. Now, you know what? In the PDF

file, can you have a indirect indirect, just like SQL injection, can you have indirect prompt injection? in a PDF document which you're storing from your your customer sent it, uploaded. Is it possible? Are you protecting it? >> So, I'm telling you these are the things which are coming in. So, this will come in and you know, will will be a big problem and we all need to

fix that. And then, the last one is the ISO. Okay, just write down ISO ISO 242001. Okay, 42001 NIST AI RAM. So, this is another one. These are the two ones which are important from security Okay? This is layer five which we are building, which is adding, you know, something called as, you know, model risk prevention. You're trying to build a model risk prevention doing that. Now,

telemetry can be added through, you know, Meteor Atlas. It's basically, you know, any of these models which are there, you know, top 10, you know, top 10 OWASP top 10 LLMs, yeah? So, those models are coming in, so you have to know all of them to solve for this. So, now there are few things coming, green ops. So, I think I had that green ops. FinOps, GreenOps,

these are also coming in. Financial operations are coming in. Why? What's the cost of is very very important also along with that. So, that said, like, you know, I think we talked a lot of good things here. This is the MITRE. You can actually go for reconnaissance to all these stages and what to do with it. Okay? That's the advantage of MITRE. uh perspective for yourself. So,

you should know this. If somebody asks you question, "Hey, how do I add security?" If you know this, you can add security to yourself when you are building this. So, real-world advisory tactics. Now, if you want to know exactly how it is used for the each one of them, you know, prompt injection, indirect prompt injection, input manipulation, you should be able to look at this in more

detail. All right? That said, you know, prompt injection is one of the big problems right now, which you need to protect yourself for uh when you're building the solutions. All right? That said, like let me just let me take a pause here for a second. Is it something you can apply to your project right away? Yeah? The goal is for you to apply this this thing in

your project. Uh you know, I had a great time talking to you guys and you know, connect with me on LinkedIn. You know, let me put this on. This is the tiny URL, which is important. Like if you want to know, this is the URL you can get more information from. I'll be available for questions uh after this session. All right, thanks, guys. Have a great time.

>> [music]