The Architecture Behind Trustworthy Enterprise AI Agents

Austin Kronz

0:00 Hi, everyone, and welcome. I see we have people trickling in still, so we'll give it about another minute or so. I am really excited for today's session. We are joined by Anu, who's the director of GenAI at Tredence, one of our amazing partners. I'm Austin Kronz, lead AI and data strategy for Atlan. And today we're gonna get practical about this, right? We're talking about how do you build an architecture that actually produces trustworthy enterprise AI agents. So as people are joining I'll go through some quick logistics. What are we gonna talk about today? I'll hand— I'll handle some of the common questions first. Everyone will ask, "Is this recorded?" Yes, we are recording this.

0:49 We will share this after the call, so please keep a lookout for that. We want this to be as interactive as possible, so please use the chat, use the comments, ask questions anytime. We'll have some dedicated time at the end of the session to handle Q&A. We're more than likely mixing questions as they come in, so if it's relevant to the conversation, we want to address it right away. But please keep the, keep the chat active, keep the questions coming. We, we appreciate the engagement. And there will be Q&A at the end and we have some folks in the chat to answer any specific questions relevant to products or anything like that.

1:25 So let's get into what are we gonna talk about today. We're gonna first start with just a quick level set on just, hey, why are agents failing without the right types of context? And then Anu is here. He's gonna talk about what Tredence has been seeing across the implementations, and then we'll get into a brief reference architecture bring this to real life, and I'll share a little bit about what this actually looks like from an AI context platform in Atlan. So before we get into it and I start talking, Anu, I just want to welcome you and, and say thank you for joining.

1:52 If you want to give a quick intro and a little bit about yourself.

Anugraha Sinha

1:56 Sure. Thank you. I'm very excited to be part of this. And, and Tredence and Atlan both have been working across the board and verticals for making sure that these confusing words, if I may so say so, what is context, what is semantic, what is knowledge base, what is knowledge graph? There are so many different kinds of words to people here today. But we have been very happy to partner with Atlan and make it real for our clients and partners across the board. And we're seeing a lot of traction as to how we take, you know, the investment that the industry has done in data and how that is being used in agents today.

2:31 And as Austin rightly said, let's dive into it. Very happy to be here with this with everybody.

Austin Kronz

2:38 Awesome. Yeah. Thanks for having me. Let's get into this. Anu, I think every conversation I have with any type of enterprise, AI leader, AI architect, something like that, starts with some type of similar problem statement, which Yeah... Listen, every single person wants to be my agentic platform. There's gonna be no shortage of platforms that'll help you develop agents.

Anugraha Sinha

3:02 Yeah.

Austin Kronz

3:03 The reality of every major organization is we're gonna have many platforms. There won't be a single platform that's building agents. There's already a heterogeneous enterprise application landscape, and so there will be thousands of agents across dozens of platforms and, you know, the unfortunate reality of that means every time you go to build an agent, it's usually siloed to whatever information or memory or context— Yeah... That that platform has been accumulating or people have built into that specific platform.

3:31 So at Atlan, essentially what we're trying to do is in the middle of this complex heterogeneous landscape on the left and this evolving landscape on the right, how do we provide an enterprise context layer that no matter where you build an agent, you have the same type of shared context within your organization?

Anugraha Sinha

3:50 Yeah, absolutely right. So because what— and, and I'll, I'll, I'll probably substantiate that point what Austin mentioned in the industry is that there are two places where these discussions happen. One is where the actual use cases lies, and they are very concentrated on the data that they want. But they kind of start losing the, the vision as to where they're building, where is it coming from, what's the larger picture. And then, then the other side of the picture is the senior executives, where they want to build this ecosystem where, you know, have the multiple different agent platforms, they have a single context layer.

4:27 And there is a void between what the reality is for, let's say, a manager who's leading a business and what data he's using because he's heavily concentrated on that. And then there is an executive who's looking at, "No, we are getting myopic. We have to think broad." And that bridge is what you see right in the cen-center over here, the mesh that gets created, and it has to make more sense for the agents to start giving more business value, but also be grounded in data and technology.

Austin Kronz

4:56 Yeah. Makes sense. I think we had talked at one point about, like, anything prefixed with the word enterprise. It's kind of it's kind of ironic because most of the time enterprise deployments start with particular use cases. And so, Yeah... You're, you're constantly, constantly playing this tug of war of we want an enterprise something platform. But we need to be able to validate it or implement it in some type of incremental fashion. So I, I do think conceptually executives understand the enterprise, like we should have some type of consistency, but we don't want to sacrifice any type of speed, especially now where— Yeah...

5:32 The market changes every single week, so we don't have enough time to, to kind of wait and do these long development life cycles.

Anugraha Sinha

5:40 Yeah.

Austin Kronz

5:41 You and I have already mentioned context a number of times.

Anugraha Sinha

5:44 Mm-hmm.

Austin Kronz

5:46 And I just want to put this out there for the audience. I added this the last minute. Feel free to scan this or go to Slido and just add the code, but I think sometimes we talk past one another and, you know, I'm saying context. Anu, you might be saying context. I'm sure everyone else on this session could be saying it, but what, what does it actually mean? So I'd love for you to either scan the code or just give me a quick like, "Oh, when I'm talking about context for AI agents, this is exactly what I, what I mean." And they should come in live and just make some type of word cloud, so.

Anugraha Sinha

6:18 Oh, that's good. That's, that's nice.

Austin Kronz

6:21 Yeah. We have one brave first participant here. Grounding.

Anugraha Sinha

6:30 In the world agents of the firm. Fair enough.

Austin Kronz

6:33 Yeah. Ground reality.

Anugraha Sinha

6:36 Ground reality.

Austin Kronz

6:37 These are actually more philosophical than I would've thought. I, The, the, the buzzwords that... It's usually buzzword soup, right? It's like I mean ontology Yeah... The most common one is probably semantic layer, right? Like, like we need semantics for our AI agents. And then if you talk to an AI builder, it's skills and tools Knowledge graphs.

Anugraha Sinha

6:59 That— Source Yeah, knowledge graph.

Austin Kronz

7:01 Yep, absolutely.

Anugraha Sinha

7:04 Maps of your business.

Austin Kronz

7:06 I'm gonna take this as a positive sign, 'cause I think this is actually very... It is not buzzword-driven. Voyage.

Anugraha Sinha

7:13 Yeah, I mean wow. Oh, this is actually advanced. Interesting. Yes.

Austin Kronz

7:21 Yeah. Cool. Well this will stay live, so I'll be curious to come back to this at the end as more people, Yeah... As more people submit responses. The, the reason I ask this is because I started to talk more in terms of a framework rather than, than buzzwords. And so at Atlan, we've kind of been using this, like, three-pronged approach, just what is context, rather than saying any one specific thing. Context is just knowledge. It's the map of your business. It's what do things actually mean. It's the definitions, the entities, how things relate. Expertise is that procedural knowledge. What are the processes, the playbooks, the workflows?

8:00 This includes everyone wants to talk about tacit knowledge, right? This includes the things that no one's ever written down. You may still have procedures that aren't just codified or, or documented just yet. And then norms, which is just what are the acceptable actions an agent can take. I think these three buckets, we can all agree that these are the ingredients that an agent needs to be successful and accurate and complete the tasks effectively that we're assigning it. So I always start with this.

8:27 The way this manifests itself and what we've seen in, with a lot of customers, regardless of the agent you're ty-trying to build, I'll give an example of an operational one and then an analytical agent, because I think those are the two broad categories right now. Let's take like a, a telco example where, you know, it's a contact center agent. Someone's submitting a question to their contact center agent saying, "Hey, I, I got this forty dollar roaming charge. Is it refundable? My service was down. I should get a refund. It wasn't my fault." I mean, even within some of these questions, there's so much of that knowledge, expertise, and norms that is baked into this.

9:04 First of all, I need the knowledge of who is this person, right? Is this, oh, this is Dana submitting a question. Well, what is roaming charge? The agent now has to understand roaming and charge are specific things in our organization, and how does a charge relate to the person? You're asking a question about refundable. Well, what are the standard operating procedures for something being refundable? You know, how do we execute this? You know, a lot of just in a simple, relatively simple question, you know, you need that knowledge, expertise, and norm for this type of question. Even in an analytical, like let's say we're retroactively trying to do this as a management at the telco.

9:46 I want to understand, hey, why are refund rates spiking in the north region this quarter? Again, within that, you get into these why questions. I think what questions are relatively straightforward at this point, just like basic sums and aggregations. But when you get into more of like a root cause, you have to think through multiple types of or multiple layers of, you know, problem and, and like key driver type analysis. Now you need to understand the semantics of the business. Well, where are refunds rates, refund rates? Is that a table in our organization? What's the column? You're saying north region. Like, have we already defined region as a column? Is north one of the attributes?

10:26 This quarter, well, is this quarter like quarter to date? Is this what we're, are we projecting for the end of the quarter? Do we go by fiscal or calendar year? These are all things that might be implied by a good analyst, but you have to more purposely engineer some of this context for the agent in and of itself.

Austin Kronz

10:49 A-as we've seen customers develop context and start to deploy context over the past couple of years, I think there's a couple challenges that they seem to run into. Everyone at some point will hit one of these three walls. The first one is we, we need to get there in a relatively k-quick manner. So time to value is really important, and context bootstrapping and starting the initial corpus of context that the agent needs to, to get up and running can't take six months. I work with a customer, they're actually one of Databricks' largest I believe they're one of the largest Genie users, and this was their challenge.

11:23 You know, they have thousands of Genie rooms that they're building. They have thousands of agents that they're producing, and most of them sort of fizzle out after a very brief pilot period because, you know, the context just isn't there. And honestly, it could take months if you really invested the time to deconstruct what already exists from dashboards and human knowledge and things like this. So we have to expedite the bootstrapping phase of this, and this is kind of where Atlan has built their platform around these problem statements. This next one is around life cycle management.

11:53 So I think people realize once you have produced something that gets reasonably accurate from an agent perspective your business changes, which means your skills need to change, which means your SOPs will change, which means, you know, there's this constant life cycle and tweaking that needs to happen, much like code, it follows a life cycle development. Context in and of itself will follow its own life cycle development development life cycle. And the, the sprawl piece starts to happen when you have your, you know— A, a much more complex agent landscape.

12:27 You know, your first agent might work well, and then you go to build your tenth agent, and you wish you could just use the things you did in the first one. But if you built your first one in a particular platform, and then let's say your next one is more of a, like agnostic harness or something like that, oftentimes you either are copying and pasting things, and that's not a great management philosophy or, you know, you're starting completely from scratch. And so you get this kind of context sprawl.

12:52 And now we're seeing, much like we had in the data world, like there there's context that is applicable for the entire enterprise, and then there's context that's only applicable for an organization, that's only applicable for a team. And they might deviate from the enterprise context slightly, but it's true. Like, you know even in our business, right, or, you know, any type of SaaS business, as you change models, maybe you flipping to more of a consumption-based model. Well, the definition that you had for your revenue is actually different. So you have two revenue metrics that are conflicting, but they're correct because they represent a different change or phase in the business.

13:29 So I'll just kinda pause here. Anu, I'd love your thoughts on this and see if this Yeah... What you're seeing as well.

Anugraha Sinha

13:36 No, absolutely. You're absolutely right. So because what we have seen, and, and very much in line with what, Austin, what you were mentioning was enterprises over the past, let's say, a decade, decade and a half more, have invested into building good data models for their businesses, which you talk about their enterprise data warehouses, their medallion architectures, and so on and so forth. And the way that data was being consumed up till very recently was that businesses used to come up, and they used to build dashboard, and the dashboard used to be the source of truth for decision-making.

14:12 Now, when you look at this, and we'll talk more about this as we go forward in the webinar, is that this aspect of how today your business is running and how is your business responding to different disruptions, which is respond run, respond, and transform aspect, which I also talked about in our post on LinkedIn, is this particular area, there are a lot of other information that comes in. For example, as you said, revenue. You mentioned in your example, what is the definition of quarter? There are, there are different fiscal years that different companies follow.

14:45 And while the word fiscal year becomes a common terminology that people use in their daily operations, in their daily communication, but from an agent's perspective, because it is somewhere programmatic in nature, fiscal year would be January to December. But you have to make it understand that fiscal year in our context is probably, let's say, from April to March, depending on what business you are. For example, if it's a BFSI organization, they might have something else. If it's a, let's say, a store which is doing furnishing, they might have a different you know, understanding to it.

15:18 And then, as you rightly said, beyond that, all of this information, when you build your first two or three agents, you can have much more control over it because then you're just looking at that one agent or one... That agent ecosystem, you just focus on one use case. Then you go ahead and build different other business functions coming into pictures, and you've got conflicting information or definitions coming in. And as you go and bring in more and more businesses starting using the agentic workflows, these definitions will start getting confusing, and then that is where the sprawl comes in.

15:51 Because now, while you have a very good layer which defines this, your agents do not know which to trust. We had a customer based out of in, in UK, and when they started building up their so-called context layer, they gave in a lot of information available on SharePoint. Now, for a specific standard operating procedure, there was a version available way back from two thousand and eight, and while the context layer just got built up, they asked a question and it gave a reference for it. Now, from an, from, from a technical person, that was the right answer, right? Nobody... Good question. Yes, there is a reference available.

16:26 But we have to make sure that this changes over time. It's a living, breathing aspect in your, in your ecosystem. It's the way your users are today.

Austin Kronz

16:38 Yeah. Makes sense. The, the context— I, I think there's a, there's a question on the chat, and I think let's probably ask them.

Anugraha Sinha

16:44 I think this is probably the second time I'm seeing this. It says that, "Can you give— get some clarity on semantic layer and the context layer difference?" If you, if you wanna talk about that, or I can take the first shot at it as well.

Austin Kronz

16:57 Yeah. Let me let me actually get into how we view the context platform, because I think it'll, it'll, it'll answer that. Honestly, earlier this summer, I, I was kind of using the old like squares and rectangle analogy to describe this a lot, which was like all, all squares are rectangles, not all rectangles are squares kind of things.

Austin Kronz

17:20 It's like all semantics are context, but not all context are semantics.

Anugraha Sinha

17:24 Semantics.

Austin Kronz

17:26 Yeah. The, the way we think of the context platform, so this middle box here is generally speaking how Atlan is approaching what is an AI context platform. And by the way, this is any... What we feel any context platform should be. Semantics are absolutely a, a, a part of the, the context layer. There are different types of context to answer different pieces of those questions that I was referring to earlier.

17:51 So I think fundamentally, if you're going to ask questions that involve your enterprise data, you will need what we would consider AI-ready data, which is Much like a human would have to have well-described, well-owned, certified, up-to-date, you know like core fundamentals of what we've built over the past, you know, seven years at Atlan, like end-to-end lineage across your source systems into your data warehouse, into your BI tools. Like anything that would help a human navigate your existing data estate in a well-described way is, is extremely important.

18:25 We've even added this concept of like SQL intelligence, which is there's actually so much valuable information if in understanding what questions are already being asked of your data, and all of that information is available. So all of the query history that's hitting your data warehouse, that can actually be reverse constructed into business questions to say over and over every month or every quarter, I can basically reverse construct all of your reporting and, and all of the business questions that are hitting your underlying data sources. So AI-ready data is still a, a very important piece. It was for humans, and now it's even exponentially more so important for agents because they'll need that as well.

19:02 So great descriptions, rich README files. How does your physical data model connect to business concepts in a glossary? You know, all of this is involved there. The semantics and ontology piece. Semantics start to be... Yes, the semantic model in and of itself is how you would define your metrics, dimensions, measures, aggregations, a-and so on and so forth. I think what, what we have found, and I've had a lot of conversations about semantic layers recently, and there is a bit of a separation between the semantic model and the semantic layer. I think semantic models with open source frameworks like it was OSI, and now it's Apache OC.

19:40 Like there's some general standardization on what a semantic model format can look like in YAML or something like that. And so if you have AI-ready data, it's actually relatively easy for our platform at Atlan to pull together a semantic model that you could then leverage for the AI agent. Now, I think where people start to blend semantic layer and semantic model is any BI platform that you have today, and now companies like Snowflake and Databricks will have their own semantic execution engine. Like you'll have a semantic view in Snowflake, you'll have a metric view inside of Databricks. You'll have a LookML model for Looker.

20:19 You'll have, you know, something else for some, and I think that is where the layer and the model concepts start to come into play more. And I'll, I'll pause there and talk more about agent skills in a little bit. But just to wrap on that, what we feel is we're taking inputs from what we consider context mining all of these systems that you have on the left here.

20:40 So whether it's your systems of data, systems of work, or systems of semantics like your BI platforms your processes, anything that you might have, we're utilizing that to then assemble and engineer context in the AI platform, and then exposing that in any type of AI platform that you would need. So whether it's your knowledge and productivity agents in your local harnesses, if it's your BI tools, your data warehouse talk to data agents or your more like operational or verticalized agents that you might just be purchasing off the shelf or, or building internally. So I'll pause there.

Anugraha Sinha

21:15 Yeah, I've shown a couple of things, so what do you think? No, absolutely. This is So, and I, I like where you start from, the, the square versus rectangle analogy, which is what. And, and, and I think that is, that is very valuable, and it kind of addresses the two questions that we see in the chat as well, which is semantic layer versus context layer. And then, then Rohit also asked you know, what's the difference between ontology and context layer? So what we see over here is a collection on the slide.

21:44 That means they are data-ready, but I think a lot of our folks on the call would have seen some layered diagrams, the cake diagrams of where the data is and what, what lives on the top of it. So I would, I would probably describe that very similar to what you have. What you see on the left side, as Austin was saying, that is your source data systems, which would be your column and rows data and the data models and so on and so forth. Goes back to your data governance, data lineages data catalogs that you have built on top of it, which is your AI AI-ready data box that you have.

22:15 On top of this is where you will map all this information. And when I say map, you're not copying data, you're building a reference over here. And I'll go ahead and spill a little bit of bean beans over here. The next layer, what we're talking about over here is assume that it's a graph that maps out different nodes, saying that you've got a table, and then you've got certain sub-nodes which are their column information. And when I say semantic, the information that this node has is also saved in the form of an embedding.

22:46 That means you can refer to it, you can do a similarity over it, and that is where the semantic aspect comes into. The word semantic is this is where because you are bringing an information compression aspect as to what that node really means. Beyond this, on top of this, if I say this is a graph, is where you'll have your ontologies built in. Your ontologies would bring in your business processes. Now, walk— let me walk you through this. Let's say there is a store in Florida which is doing business, and it says that my store's sales are affected by weather. Now, these are two specific nodes in the system.

23:26 You've got sales data, and you've got weather information, and there is a direct relationship that we can create in a graph that sales is affected by the weather. However, this relationship that you have on top of it can have a very store-specific weightage. The store in Florida will have a very strong weightage that weather affects my sales, but maybe if you go somewhere in Midwest, they'll say, "No, it doesn't really affect mine." That is where your business aspect will start influencing these relationship weightages, and that is where your ontology comes into picture because the business operation is saying, "Yes, weather does affect me." Somebody's gonna go ahead and say that, "Yes, it affects me more."

24:05 Somebody's gonna say, "It affects me less," and that is based on the operations your other business managers are doing in those areas. We Now you've got an ontology layer, you've got references from an embedding perspective of these nodes, you've got relationships, but your data continues to remain where it was. No change to that. You're not copying data altogether. Then on top of it is the access part of it. That is gonna come in. And what we have done in this layered and this analogy is, your bottommost layer was your EDWs, your columnar data. Then you got relationships built up.

24:38 Relationships also reference through an embedding so that we can do some comparison for natural language queries, and ontology is also brought together. And this all together will become your context layer. Combined information will become your context layer. So you've got context which which— and going back to where we started from. The bigger box was a smaller box, rectangles and squares. Context layer is a collection of things. You've got ontologies, you've got knowledge layer, knowledge graph or knowledge layers, and then you've got your actual EDW, and business operations also reflected in the form of ontologies.

25:10 That's the stack that you look at which starts— which is being consumed by agents for giving answers, replies, things of that sort. Skills is what the agent would use. Like, for example, I wanna, I, I wanna have a skill as to how do you read sales information. It is gonna use different aspects of your context to give answers back. But that's what I would wanna add on top of what you see on the screen. Back to you, Austin.

Austin Kronz

25:35 Yeah. And I, I think so if anyone's familiar with Atlan, right? We have— we had a long history. In fact, I found Atlan when I was a Gartner analyst still back in twenty nineteen was I think the first time Prukalpa, our co-founder, had shown me Atlan. And even back then, Atlan's mission was to be the context layer for data teams. So pre-AI, pre-gen AI, this was like it was always like, "Hey, how can we add context for the humans of data?"

26:05 What that had allowed us to do for a, a quite a long time, like coming up to twenty twenty-three when ChatGPT became, you know, a thing, was we've had a lot of experience building rich metadata connections into these enterprise systems. And so whenever we released, we were one of the first companies to release an AI product back in twenty twenty-three which means we've had a lot of learnings with customers on how important context is in terms of getting something into production or getting something into production in our own platform. These rich connections, you were talking a lot about ontology, and I liked your weather and store example.

26:38 Some people ask us, "Oh, do you build the ontology?" Or, like, "What if I have my own ontology?" And, and of course, the answer is, if you have it, great, we'll start from that. What we found is there's so much rich information in even your structured data. So think of, think of, you know, a, a, like how does weather affect store performance? More than likely, an analyst somewhere has probably written a query that says, that joins information from weather conditions to your sales table, right? And so there might be a recurring query that every summer during storm season, we're getting the same type of question over and over from a SQL statement.

27:15 So even in that sense, like relationships can be captured and relationships can be inferred from some type of structured data. So when we talk about like context mining and what Atlan's doing, a lot of it is all this rich information that already exists across your key systems. Like how much can we bootstrap from that? And then layer on, like, hey, you've never documented that this is a repeated analysis that you do, and like maybe we just need prompted to create that recurring skill or that procedural knowledge. But that's all things that we can kind of evolve and iterate with as you, like, hill climb towards the agent performance.

27:50 But and maybe just one last thing, because I see there's still, there's always gonna be questions about ontologies and, and I, I saw you responded in there. All of these are important, right? Sometimes I just like to level the, the conversation and say, can we agree that relationships are very important for agents? Relationships can come in the form of ontologies, nodes and edges in a graph. They can come in the form of primary foreign key relationships and structured data. They can come in the form of wiki, wiki memory systems with text-based files. They can come— There's all these types of things.

28:29 So within this context platform, I think it's good to call out that whether it's keyword search over definitions, whether it's—. -a vector search over, you know, a semantic search over a vector database, if it's a graph retrieval, if it's a semantic retrieval, like whatever that— It's going to be a hybrid of all of those things. This is what a context platform should, should address is like how can we effectively navigate relationships to give agents the right context?

Anugraha Sinha

28:59 Ab-absolutely. And I think I, I kind of like the kind of questions that we're seeing on the chat which which, which gives an impression of the kind of audience that we have. Very interesting questions. One that just came in from Surya is Surya, I would that's a very good question about SQL query and the right way to generate sequels. And I know we started with text to sequels the world probably about eighteen months back, and now we have morphed into a much complex aspect. That's a detailed question. I'll try and address that question when we discuss the architecture, Surya.

29:29 I hope I'll be able to answer your question, but very happy to connect otherwise as well.

Anugraha Sinha

29:33 Just drop me a message and then we can go deeper into it. But that's a very, very important question. You I can probably do some kind of a VS Code session along with you to see how that works out and separate to this one that we have. But moving on to the next section what we have would— And, and let's go to the next slide. That's absolutely fine. What we are seeing in the industry, which we talked about earlier as well, and and I'll There are two ways to address this. We can start ground up, and I see the, a lot of questions in the chat as ground up.

30:03 That means starting from data, where is your context layer, where is... And we are not talking about the business aspect of things. While some of our folks in the audience would be business people who are taking business decisions and they are talking from a top to bottom. Let's, let's swift, let's switch and let's talk top to bottom right now. If you look it from your enterprise leader's perspective, when they look at building agents or agentifying, quote-unquote the business operations, there are three main things that comes into picture. One, you have your daily operations that your company does, which is your how do you run your business.

30:41 The second is every single business, no matter how stable, how unstable they may be, they are susceptible to disruptions, and you respond to those disruptions. In the, in, in, in the area of retail, there is supply chain disruptions, there are geopolitical disruptions. You've got disruption with respect to promos, competition, events, so on and so forth. Last but not the least, your enterprise leaders are also looking, continuously looking at transforming their business. When you're running your business, if you go to any particular business manager, they've got a standard set of operations that they do in a store, for that matter. Consider any grocery store, for that matter.

31:23 I mean, the one the biggest in the US right now, Walmart. If you go to a store manager who's putting up requirements going back to their warehouse, every single day he comes in, he goes to his office, spends two hours over there looking at the data that is there. Then he'll come down on shop floor to start talking to people. He spends those first two hours over there. That is the way he runs his business. Then comes response. There is weather disruption that comes in for that store. Again, going back to that grocery store. And how do you respond to it? There is a competitive store that comes out in the nearby.

31:57 How do you go ahead and manage your pricing, your promos, and so on and so forth? Last but not the least, can you go ahead and connect across different business functions together and try and come up with a new strategy? Can you do a cross-business function query and let the agent do a root cause analysis across multiple multiple different w-business functions and give response back? What we did, and what you see over here as a case study, is that we were working with a Fortune five hundred apparel retailer, and they have stores across the board. Not in US, but across the world, and multi-multiple different brands under them.

32:35 And they are they already have invested a lot in building their enterprise data models over the past all these years, and they've got great, you know, dashboards and some semantic layers and business KPIs, definitions built out as a function of those dashboards. And And they're, they are taking decisions today. But then again, those two hours in the morning is where the guy who's sit-sitting down and taking decisions or planning is spending.

32:59 What we have transformed from them is, every morning when the guy starts and sits in his car to come to the store, he gets a report on his phone, and he can press a speaker button over there, and the report gets read out for him. What happened the past twenty-four hours for every sing-sing-single business. The moment before he parks his car outside the store, he already has a lot of information as to what has happened.

33:24 And that, those agents spin up at two o'clock in the night for every particular store, do a wide spectrum analysis about what product, which segment, how it happened, generates a report, sends across the email, generates an audio, and pushes to the mobile device. And he can understand, go down, and the difference that is bringing in is, now that person, when he goes to the office, he's spending only twenty minutes in the office to go through that report, and he's seeing the report in the same manner every single day. So he's getting used to it. It's not it's not difficult for him to say the same thing again, understand same thing again every, every single morning.

33:59 It's the same thing, similar format. And he is back on the floor talking to his floor agents to take faster decisions. And that is where business ROI is coming into picture. What we have seen is that by building that context layer, we are building these agents about sixty percent faster. They started with their first two agents in about six, in about twelve weeks, and now they're rolling out agents every third week. Because now the context is being reused again and again because those nodes, those definitions, those business operations have now become much more easy for agents to access, and they don't have to reinvent the wheel again singl— at every single point of time.

34:35 And the recommendation accuracies are also getting better. So we'll, we'll, we'll get to that, Shubhangi, on the structured diagram, which is a couple of slides beyond this one. And this aspect as to the last part is, you know, the agent starts to reason the way the business actually operates and not some kind of an elevated thought process which leads to certain hallucinations that, and today you ask a question, the agent is gonna give you a response, and you ask the same question again, and the response is totally different. It looked up something very, very different because it had so many options to choose from.

35:08 And that is where your context layer and semantic aspect and knowledge of business operations built out on ontologies make a hell of a difference. Let's go to the next one, Austin. And I'm looking at the questions. I have a the pane open. I'll, I'll address some of the questions as I go forward. Very similar, but a different problem statement. Healthcare industry. Extremely heavily regulated. Manufacturing lifecycle. Manufacturing processes. Now, you can understand when we go ahead and buy the simplest of medicine from over the counter, we expect the quality of every single capsule, every single drop of a cough syrup to be as effective as the last one or as effective as every single molecule.

35:53 You would imagine that the amount of testing that happens on this needs a human interface, needs a human validation, irrespective, and obviously statistical analysis. A hell amount of time gets spent in evaluating those batches every single point of time. What we did was, and the regulations around this are pretty strict, but they're well documented because the same regulation applies to every single organizations.

36:17 We, we were able to go ahead and implement an agentic process of looking at the test results Looking at the past tests that have been done where things have gone wrong, and the and the guidelines and the frameworks brought up by, you know, the regulatory authorities, and give a quicker mechanism to batch evaluation people in the on, on a manufacturing floor take faster decisions. But the most important thing is the human is still the person taking the decision, but his decision-making power is way more faster. And why does that really help? And I'll give you an example on the same lines.

36:54 If one person is evaluating batches, he's taking X amount, X amount of minutes to go ahead and evaluate that batch, and there's a lot of backlog before that batch goes out on the truck and to the warehouse. Now they have a better way of consuming that information with a much structured manner, repeatable reports which they can look at, and recommendation coming where something might have gone wrong. The person is still doing the same job. He's not changing, but he's doing a job much more faster because he's not doing that same operation again and again.

37:21 And that is where we have followed fifty percent faster batch release cycle times in this particular use case, and the traceability is definitely there because why was a particular recommendation given? What was the result? A clear-cut reference is given to them because you have a very solid context layer. Because I know what reference did I pick up from the graph, from the ontology, and why that inference was made.

37:46 So again, another example of how enterprise context layer and what we talked about earlier, as Austin mentioned, the different layers over there, different boxes over there, if you have that well-structured, built out, and then you have the agents talking to it, it gives a lot and lot of benefit in the longer run. We have some more questions over here, and I know people have been waiting for the architecture as well. We'll go down over there in a while. But Austin, any thoughts on the past two things that I talked about on the use cases or, or what you are seeing in the industry today?

Austin Kronz

38:22 Yeah, I think I, I just wanted to call out you know, everyone can see the, the clock, but I think we've been keeping up with a lot of the questions, so I think we're, we're in good shape here.

Anugraha Sinha

38:31 Yeah.

Austin Kronz

38:31 There's, there's been a lot of questions about harvesting or how did you get these SOP... How did you get you know, like some of this, this context? And I think it's, it's worth commenting on. I had a slide about kind of like the context assembly, context enginee— context engineering process. Maybe I'll try to summarize this and I'll see, I'll see what you think, Anu.

Anugraha Sinha

38:53 Yeah...

Austin Kronz

38:53 From a context platform perspective, obviously, we would love to default and automate as much as possible in terms of harvesting context. Now, it's relatively easy, relatively, I shouldn't say easy. What we've invested in is a custom harness to extract context from your underlying system. So whether that's your databases, data warehouse, all these, all these tools, that, that can really help us build a solid foundation in terms of how you describe your data, your assets, how you build out semantics, how you build even some foundational skills we can help with. There, you know, can help bootstrap some of the ontology. There's a lot that we can bootstrap.

39:30 There's always gonna be an element of, okay, some of this now lives in our documentation. It doesn't live in the code of our systems. And so, yes, maybe in a local harness, maybe in something that you're building, you're gonna go through some type of context assembly process and just relevant to Atlan, storing knowledge, meaning unstructured data. There's a concept of knowledge fol-folders and files where, like, if you do have actual SOP docs or you have a Confluence page or you have something that you might want to incorporate as context for an agent, you can absolutely do that.

40:02 To your point, like, you might be pulling from something from two thousand and eight, so there's a whole life cycle around that as well. An agent doesn't need your twenty-page Confluence doc on how you do customer success at your company. It might need the twenty rules and just codify that into some type of either skill or just knowledge file, as we would call it. So there's context assem— there's automation, there's context assembly, and then the number one question is, like, tacit knowledge. Like, we, we don't have any documentation. We wish we did. And I think there's a lot of posts around, like, how you can do that.

40:35 Some people have done interviews and recorded them and turned that into documentation. Some people have, like, I think Anthropic literally can screen record as you go about your day and try to reverse engineer skills from it. Like, like, I'm not gonna sit here and tell you there's a magic button that can somehow extrapolate all context. Bootstrapping context is one of those three avenues, and then the holy grail of, I think, agent improvement is, are, are traces. So how do, how does your consumer actually engage with the agent that you produce? Hopefully, we've given it enough context to start out with that it's getting some percentage of the questions correct.

41:16 But when you incorrectly utilize a skill or a tool or there's a gap in knowledge, how can the traces then feed back into that context development process and actually more dynamically improve that? So I just wanted to, I just wanted to call out because there's a lot of questions about assembly before we get into the, the architecture, and I know we only have four minutes left, so we might have to skip the demo on this one. But happy Yeah.

Anugraha Sinha

41:39 Let's go to the architecture because I think a lot of people have been talking about, on the chat about architecture. And I, I, I, I wanted to be sure that I do not have too much information out over here, but then I wanted to be sure that I get all the layers because some people talked about auth in the ecosystem.

Anugraha Sinha

41:59 People also talked about where do the agents live. So let me walk you through this, and I'll probably— Under no pressure.

Austin Kronz

42:05 You have, you have three minutes to go through sixty— I know.... Pages of content, so.

Anugraha Sinha

42:11 Thanks a lot. So I'll quickly talk about the top right box. You would see that you've got multiple data sources. You had these PDF document, Confluence pages, SharePoint locations, so on and so forth. All of that data and your transformation and gold layers, that what you have is your medallion architectures, the bronze, silver, gold. From there, you go ahead and build your dashboards, which is the de facto decision-making ecosystem up till now, if I may say so. And as I mentioned, all of this information goes into your knowledge graph. Then is where your ontologies and metadata gets built in.

42:46 Those relationships, those, those business aspects, acronyms, business definitions, specific KPI calculations, and all of that comes into picture. And the vector databases that the, the unstructured data through your document processing pipeline over the vector database also gets represented at the knowledge graph and the ontology on top of it. That means for the agents so to say, these graphs and ontologies becomes a source of truth as to what what do you need to investigate under what kind of question. That does not discount the fact, and which is one of the questions that we had on the chat, is the structured data aspect.

43:25 So when you do go ahead and look at what information needs to be extracted for answering this question or this operation, you do go ahead and at at one point of time, some kind of SQL query will have to be created. And if it's a reference document, you'll have to bring it in the chunk right into the prompt so that it can make sense of it and then or it could plan the execution of skills based on those chunks that have been provided as a reference. But then you would have to create a SQL query to execute because when you're looking at the data.

43:53 But which data to look at, which table to look at is the information that you can shortlist from the other ecosystem. The auth aspect that you see is at the bottom of the aspect, and some people did ask about how do you control it. There are essentially four specific design patterns as to how you transfer authorization for data. That means who's asking a question and how do you access that. We can deep dive into it, of those design patterns, but it is a foundational element. Agents will have an identity, but they will borrow the identity if it's an agent that is being used by users.

44:26 And if it's an autonomous agent, you need to control the roles and responsibilities that that agent has based on service principle and service accounts so that they can access the right data. It's a deeper topic, but authorization is definitely a backbone in this. That means not all agents will get access to all the data, which is essentially a bad design. You have to control that. I, I have— I'm probably out of time already. I'll take a pause, and the slides will be available. Very happy to have this discussion with anybody else otherwise as well.

Austin Kronz

44:52 Anu, what— do you, do you have a call right now?

Anugraha Sinha

44:56 I can extend for sure. I can extend for some time if people wanna stay up. I'm absolutely fine with that too.

Austin Kronz

45:01 I think I'm gonna make the executive decision. We're going five more minutes.

Anugraha Sinha

45:04 All right.

Austin Kronz

45:05 If, if if people have to drop, we will send the recording. But I, I think this is really helpful. So let's, let's do five more, and then we can wrap up.

Anugraha Sinha

45:13 Sure. So I talked about the top right corner over here, which is third-party data sources and your enterprise data getting mapped into a context layer. And because this is more on the context side, but just to walk you through as to what this all is. You will have your business applications, calling applications, whatever it what, whatever and whatnot. You will have an API management there, which is the front gate for any questions or queries coming in from people. You would you would need to go ahead and and you would assume that there is some kind of an application gateway with some bookkeeping the data for your— these applications.

45:47 But then you have an AI gateway and your MCP toolkits. That means these MCP toolkits are the gateway for the agents through the gateway, through the AI gateway to access the data. Why do we need MCPs? Because it's easier to implement it rather than you asking the agent to build a SQL query and then execute it. You can probably go to a MCP server to execute that SQL. It's easier, it's reusable. And then another aspect is that your gateway would also give access to these model gardens of the world that you might want to... And sometimes you're using this model and that model. But again, the gateway sits in the center of it.

46:23 And when I'm saying the gateway, it's the HTTP REST API gateway that you see in, in the center, not the API management. This is an AI gateway, not an API gateway. An AI gateway is very specific to giving access to AI ecosystems, models, MCP servers. Why? Because you need to control the authorization at that layer. And it integrates with your auth layer naturally. It's not an exception. Over the agentic ecosystem, you need to have certain question banks because it's very important that when, when a question comes in, you don't start planning at every single query how that question needs to be answered.

46:57 That is unnecessarily gonna slow down things, and hence the task planning templates makes a lot of difference for any of the other aspects, you know, intent parsing, semantic routing, insight agents that need to be built in and connected to MCP servers, and the other skills that you want to bring in, which makes, which makes sure that these agents should have access to some of your AI ML capabilities, data science models that you may have. Report generation capabilities may require certain code generations aspect and other other ways. What are task planning templates, Hiren? When you take a question, you need to map how that question needs to be answered.

47:31 And why that is important is because your users will trust your application if you give them similar answers for same questions every point of time. So if you have templatized way of answering certain types of questions, it will execute those skills in the same manner, whether parallel, whether sequenced, whatever that might be, so that your end application has a repeatability of the answers and responses that are given in. Maximum another maximum question. Gartner refers to context graphs to capture decisions, exceptions, logics, yes. Is this context graph should be relate to long and short-term memory in this figure? Yes.

Austin Kronz

48:10 I love it.

Anugraha Sinha

48:11 So as Yeah. Yeah, that's, that's a very good question. When we say context, we've all talked about context till now with the fact what your data is and how you're building context. But, and good that you mentioned this. This context layer that you build in, it's a breathing, living, breathing ecosystem.

48:29 That means from a short-term perspective and long-term perspective, your agents, what they have done and the feedback that they get from the user, they need to maintain that context from a short-term perspective in that particular conversation, so, so to say, and also a long-term perspective to look back And say, you know, not only in this conversation ID that I have, but you would want to look back and say, what are the longer term aspect of how this needs to be answered? But Austin, please go ahead if you want to add more information.

Austin Kronz

48:55 Yeah. No, I love, I love the question. I agree with everything you said, Anu. Definitely related to, to memory. In fact, the idea of a, like, a context graph to me is to capture edge case decisions that your existing context couldn't, couldn't account for. So in that telco example that I just gave in the beginning, like if I'm getting a, a refund, let's just say I, I built the system to allow my agent to automatically give refunds as long as it's before, you know, midnight on a particular day in their plan or their contract. But I built that based on East Coast time or East Coast is somehow in my system.

49:35 And it was eleven fifty-eight in Pacific Time or, you know, and they're, or they're in Europe or something like that. And so theoretically, I just built the system wrong, and it has, it is a true edge case where this person should get a refund. The agent should be able to capture that, okay, the human overrode the decision and actually gave this refund, and they've done it three times. And so to me, the active context graph is the living, breathing piece where traces are actually influencing the context layer. And now that we've learned this is a repeated behavior, yes, it absolutely should work its way back into a standard procedure or ultimately just fix the system.

50:14 Like, like that was a, a poor rule in and of itself. So yes, it's kind of like extrapolating learnings, not just from not just— it's, it's updating long-term memory in many cases based on actions.

Anugraha Sinha

50:29 Austin, you took an executive dec-decision earlier. I think there's a call from the audience saying that they would want to see a demo from Atlan as well. I'll give it back to you. You again be the executive over here. And I'm sure five minutes is gonna be not a just— it would not do justice to the demo that we can do, but you go ahead and take the call.

Austin Kronz

50:48 I, I think what, what I'll do is when we send the recording of this, we'll send the product walkthrough of the context layer in Atlan— Yeah. Because it's very similar to, to what I would share. I see a couple people asking for... I, I, I can do five. I could do a couple, a couple minutes on this. Yeah. The, the demo that we'll send is a recorded demo. I think that'll just be, be easiest. Well, if we're gonna use the full, let's just... Anu, should we use the full hour?

Anugraha Sinha

51:20 I have no problems, man. Seriously, I'm I'm loving this discussion and, and the kind of interactions, questions that we're getting. I could push off some of the other meeting. This is tremendously amazing conversations.

Austin Kronz

51:31 All right. I have one question for you before I get into an Atlan demo.

Anugraha Sinha

51:35 Yeah.

Austin Kronz

51:35 So this is an extremely comprehensive, I-I'll, I'll say build, but that might not even be accurate diagram.

Anugraha Sinha

51:45 Yeah.

Austin Kronz

51:46 Even in the, even in the first half of this year, how have you seen components of this change? Like this, this looks kind of daunting. Like I'm just trying to build my first handful of production agents. Is this, I need to do all of this just to get an agent to work? Like, like how would you advise that kind of crawl, walk, run that you talked about— Yeah, yeah. -before- Right. -within this diagram?

Anugraha Sinha

52:08 Absolutely. The reason, and then when I was building this out, I had in my mind that can I just go ahead and build simple boxes and not have so many articulate? The reason why I took the second part is because the simple boxes could be high level, but I wanted to make it as detailed as possible. Now, to add to the question, how are we seeing the evolution of these aspects in these modules over the past or, or in an enterprise from a crawl, walk, run perspective?

52:33 My suggestion and the way we have seen work for a lot of larger industries, bigger and small both, is that first thing is you have to understand that your bare level data, that means where your actual data lives in, there is some amount of sanctity to it. You should trust that. I have seen some of the business leaders who are honest and say that they don't trust their data there because they've not invested enough. If you start out from the context layer, but your data background is not, you know, enough sanitized, I think there you know the answer to that question, where you should be putting in your information.

53:09 And that is where the data aspect comes in from a crawl perspective. But the crawl perspective from an agentic part is you do not have to start looking at an enterprise platform on day one. Start with one particular use case and the, and the related data and context that you can bring in, and that is your, your seed for the bigger forest that you want to build in. However, while you are starting on that use case journey as your first step, start to think which service provider or cloud provider you would want to build this in.

53:39 And that is less of a technology question, honestly, to be very honest, because right now you get these capabilities out of the box from Databricks, from Snowflake, from Google, from Azure, and so on and so forth. It's more of a strategic question as to what way your organization is aligned with. Once you have that alignment AI gateways, API gateways, the ability to build agents, orchestrate the agents, observability, traces, spans, long-term, short-term memories, everything, all those are turnkey solutions today, which at least just at the end of December 2025, was being custom-built by a lot of companies.

54:15 But post January, some of the you know, announcements from Google the Google Summit in April, in April, and then the Databricks AI Summit in June, a lot of these components that you see over here are now becoming turnkey solutions. You don't have to build them. They're essentially as a service available in these ecosystems, and you can start utilizing them. That's where you move from crawl to walk. And the run aspect, it goes back to your slide, Austin, where you said you've got thirty-plus agents being built on that platform. You also reusing your context layer.

54:47 And now you can churn out agents and agentify businesses in about two to three weeks because now you have a very trusted source of information and a standardized way to deploy the agents.

Austin Kronz

54:59 That's great. Should I do a five-minute super accelerated demo?

Austin Kronz

55:06 Quick demo.

55:08 We can, we can try that. All right, let's see. Still sharing my browser. That works. So I'm gonna start not in Atlan which I, I think is apropos for where the industry is. I think you should be integrated into something. So we actually built this demo. If you hadn't seen anything we had built out over the summer at the Snowflake or Databricks conferences, we kind of built this fictitious burger franchise that we call McContext, and that's where all of our demos are, are based on. So the screen you're actually sharing, that, that you're actually seeing are just some managed agents that we have built to demo how context impacts agent performance.

55:48 And so what we put together is this basic analytics agent for a franchise operator named Marcus, who has some basic questions of his burger franchise, like, "Why is my drive-through time up this week?" And much like before, there's so many implicit understand— so much implicit understanding of parts of that question. Like, well, first of all, what is drive-through time? What is this week? Is this week Monday to Friday? Is it Sunday to Monday? So if we run this, what we've done is done a we, we've put together a side-by-side, which just says, let's just let... On the left-hand side, we'll just let the model figure it out. You know, very basic.

56:22 It's a Claude managed agent, so Claude's the hardest behind the scenes. But what will it do? We'll just let it go to the database. On the right-hand side, we'll have it actually leverage Atlan as the context layer, and we'll see the results in the bottom. I think the important thing to call out is you do get an answer. Like, if you just let the model, if you give it access to the database, like, you will get some type of answer. It's actually being a lot more honest than I, than I would expect.

56:51 I forget which model we're using for this, but it's, it's actually saying, "I don't know some things and I have some follow-up questions." It, it really wants to understand, well, who, who are my... So it doesn't really have any understanding of the persona in and of itself. It's saying, "I'm assuming something for now, but I need you to confirm or direct me." And so after, I think, multiple passes of you probably manually injecting context into this agent loop in this conversation, you'll get to some answer.

57:18 Ultimately, it did come up with, "Hey, we, we do see that over the past seven days there's It, it is, it has increased over a previous period, but, I mean, beyond that, it's not really giving you a definitive answer of why this drive-through time is up. And you can see it kind of just went directly to the database to write some queries. And it went to the raw layer, which is interesting, so hopefully that's okay, but more than likely not.

57:44 In this example with Atlan integrated as the context layer into kind of the harness what we have seen is, first and foremost, it's loading this concept of a persona, which is just a knowledge file inside of Atlan that says, "Oh, we understand who Marcus J. Is. We understand he's a store owner." Oh, we also found some terms coming from the Atlan glossary. So drive-through time actually exists in Atlan, combo revenue and its definition, operators and its definition. All of these terms actually exist inside of the context layer for Atlan.

58:15 And, you know, you're asking a why question, so we know that this would actually trigger this skill called diagnostic playbook seasonality check, just anything a good analyst would do, what ifs, because we want to know root cause, and we want to know, like, hey, what happens if, you know, we don't do anything about this? And then we get into semantic models, right? It's loading different models that are relevant based on the question and ultimately then going and executing the queries. So when I say executing the queries, Atlan is not doing this. Atlan is the reference point. This is Snowflake, Databricks.

58:48 Where-wherever you're choosing to send these, whatever your execution layer, data layer is of choice, you know, this is actually maybe Cortex running this query, or it's just raw SQL against Snowflake or Databricks or wherever, Google. And then we get a result, right? And you'll notice the, the difference here is that it's more contextually relevant. It's first, it understands that there's some type of promo. It understands that there is a relationship between promos and drive-through time, so it has some contextual understanding. It doesn't just know that there's some, you know, historical increase or, like, relative historical increase. It knows that there's some type of benchmark that you've already established.

59:27 It knows it doesn't have anything to do with seasonality because weather shouldn't affect this. And, and it not only gives you an answer to why, it also kind of goes further and says, "Hey, here is some potentially additional relevant information and the cost of doing nothing about it." So you get a much more rich response. So we just like to start with this because I think the common question is, what does this look like when I actually build an agent? And if I work my way backwards with, okay, so how did we get to the point where all these little Atlan artifacts are actually built? I can come into Atlan in and of itself.

1:00:03 Atlan has had a long history in building out, you know, your data graph, and so you'll see rich connection to, like I said, your data warehouse, your databases, your source systems, BI tools, transformation tools, et cetera. All of that is still part of our, like, AI-ready data catalog. The additional pieces that we've started to add are more of how do we help you engineer and assemble context. And so you'll see this context engineering studio where, yes, this is our own harness where we will help you assemble the context artifacts like skills, knowledge like really anything that's, needs to be called by an agent. You can do it in natural language.

1:00:42 You could do this in your own harness. Like I might do some of this in my own Claude Code or Codex or something like that, and then just use the Atlan MCP to publish it back to Atlan. As you go through and assemble this, I'll try to find an example of, like, what this agentic loop would actually look like in our harness. So I give it a... I'm trying to build this talk to data analyst about McContext franchises. It will search through everything that exists in Atlan already and realize, like, oh, you actually have some tables here that are relevant.

1:01:11 It might even come back to me and say, "Hey, these don't have really great descriptions. Do you want to use Atlan's context agents to enrich the descriptions?" And you can just say yes and keep prompting through this.

1:01:22 The fundamental unit it will start to build is like a context repository. So just like code has a GitHub repository, context will have its own type of r-repository for versioning and, and things like that as well. So you can keep prompting through this loop. It'll start to generate skills. It's very interactive. So like I said, there's no, there's not a magic button for everything. What we can do out of the box and where we've invested for the past few years are really getting to that AI-ready data point.

1:01:49 So how do we do metric conflict detect— or how do we do metric detection, metric creation, SQL history, descriptions, read-mes, things that are really rich based on the underlying data graph. And then it'll assemble itself into this kind of agnostic format. So you'll see here, yes, you're looking at a semantic model. Keep in mind, this was built for like Snowflake in this example, but this is just a complete agnostic OSI spec semantic model. And so you could deploy this wherever, and this is really the whole benefit of the enterprise context layer.

1:02:21 Yes, we have semantic models, we have skills we have different knowledge files and folders that might be a part of this, and this is all of the pieces of context that your agent will need to be successful. The benefit of doing this and having this here in, in interoperable are you might want to use skills in a new agent, in a new harness, in a new platform tomorrow, rather than building everything within that one agentic development platform. You can assemble all the context needed in Atlan and then just call it via MCP or a direct kind of publish. However you want to access it really is, is there.

1:02:56 We'll help you simulate some of these, so for analytic questions, right? Like we're already doing some of the e-eval-based questions that we've reverse engineered based on your actual query history. And obviously there's, there, there can be human in the loop here. So like I said, we automate where possible, augment the human where possible, and then anything else, yes, you might have to manually upload or curate. So this is my overly long rapid fire, but I'll, I'll kind of end with that. Anu, any, any last comments here before I jump off?

Anugraha Sinha

1:03:27 I we've, we've been taking examples, and I, I don't want to extend this further. I would like to thank everybody— Yeah. -who's joined in. Wonderful questions. This was one of one of the best webinars that I've done with so, so much amount of questions and interactions from the audience. We'll be very happy to have conversations with any one of them. Please drop in messages on LinkedIn, and we'll— me and Austin will both get back and get in touch with you for any of the other deeper discussions. Thank you for joining in. This was tremendously great. Thank you for time. Thank you for extending for everybody who has stood back.

1:04:01 And thanks, Austin and Atlan team, for organizing this.

Austin Kronz

1:04:04 Yeah.

Anugraha Sinha

1:04:05 Very happy to partner with you.

Austin Kronz

1:04:05 I plus one to everything you said. Thank you for everyone joining, sticking around for the extra time. I'm sure there's more questions now than ever, but follow up with us. We'll be in touch. You'll get the recording. And I just— Anu, I just want to say thank you to you and the Tredence team as well. I remember our first conversation with the Tredence team, working with one of the customers. When we got off the call, I, I pinged some of our team internally, and I was like, "The team that are working on agents at Tredence are one of the most, like, bleeding-edge teams that I've worked with." So shout out to you and your group.

1:04:35 I, I really enjoy every conversation. I learn something every single time I talk to you, and I look forward to the next time. So thanks, everyone.

Anugraha Sinha

1:04:41 Thank you.

Austin Kronz

1:04:42 Take care.

Anugraha Sinha

1:04:42 Thank you. Bye.