Bringing Agents to Data: Enterprise AI Architecture with Iceberg and Unity Catalog

Summary

Workday and Databricks demonstrate how bringing agents to a central data layer—rather than building point-to-point integrations—converts the multiplicative complexity of M agents connecting to N data systems into a simpler additive architecture.
Apache Iceberg provides ACID transactions and snapshot isolation that allow both rapid AI experimentation and deterministic financial processes to run on the same datasets without data integrity conflicts.
Unity Catalog acts as the unified governance layer for both agents and data, enforcing fine-grained access controls and audit trails that enable 100+ coordinated agents to operate safely without bottlenecking innovation.

Bringing Agents to Data: Enterprise AI Architecture with Iceberg and Unity Catalog

Traditional point-to-point agent architectures create data silos and governance gaps that fail at scale. Workday and Databricks present a production-grade alternative: bringing agents to data through a unified lakehouse foundation. Using Apache Iceberg for ACID consistency and Unity Catalog for centralized governance, enterprises can operate 100+ coordinated agents across departments while maintaining fine-grained security, audit trails, and real-time data integrity. Learn the architectural principles for converting complexity from a multiplicative problem into an additive one.
this video covers how Apache Iceberg enables both rapid AI experimentation and deterministic financial processes on the same data, memory management patterns for multi-agent workflows, the role of Unity Catalog in enforcing governance without bottlenecking innovation, and the upcoming storage layer enhancements (adaptive metadata, multimodal data types, column operations) that unlock agentic workloads at enterprise scale.
🤝

Chapters

FAQs

Why bring agents to data instead of copying data to each agent?

Moving data to each agent creates isolated copies, introduces governance gaps, and produces a multiplicative complexity problem where every new agent must integrate with every data system independently. By bringing agents to a central data layer built on Apache Iceberg and Unity Catalog, organizations convert the M×N problem into an M+N additive architecture that scales without exponential integration overhead.

How does Apache Iceberg support agentic AI workloads?

Apache Iceberg provides ACID transactions and snapshot isolation, which means multiple agents can read and write data concurrently without corrupting each other's view of the data. This makes it possible to run rapid AI experimentation and deterministic financial processes on the same underlying datasets simultaneously, which Workday identified as a core requirement for enterprise-scale agent deployments.

How does Unity Catalog govern AI agents at enterprise scale?

Unity Catalog applies fine-grained access controls and maintains complete audit trails for every operation an agent performs, ensuring that governance does not require manual intervention as the number of agents grows. This allows organizations to operate more than 100 coordinated agents across departments while satisfying compliance and security requirements without creating bottlenecks for engineering teams.

What storage layer enhancements are coming to support agentic workloads?

This video describes upcoming Apache Iceberg enhancements including adaptive metadata management, multimodal data type support, and new column-level operations designed to reduce the overhead of managing large agent memory stores and mixed-media data. These improvements are intended to unlock agentic workloads at enterprise scale on the Databricks Data and AI platform.

Full transcript

[00:08] Welcome everyone. It's late on a Wednesday afternoon. So you're here, you're like semi-awake. We'll try to keep you awake for the next 40 minutes. If there's any constellation, we're also running on caffeine and fumes at this point. So we'll get through it together. But I actually think this is a really interesting topic. So introductions, my name is Jason Reed.
[00:24] I'm a director of product management at Databricks. I oversee everything open lakehouse, which includes storage formats, Delta Lake and Apache Iceberg. I actually came as part of the Tabular acquisition where I was a co-founder. So Iceberg's been like in my history a long time. And then Unity Unity catalog as
[00:41] well and everything we do with interoperability across the data ecosystem, you know, catalog federation, etc. And I'm joined by Phoenix. Thank you, Jason. And good afternoon everybody. In fact, good evening almost. I am Phoenix, Phoenix Majumder.
[00:56] I'm senior director at Workday. I lead our data platforms, our AI platforms, AI engineering and data science teams. This topic particularly is very near and dear to me because when we started formulating this problem, we didn't have a textbook answer. So a lot of it was
[01:13] unknown. And navigating through the unknown what makes the journey fun. Yeah, absolutely. So just some like context on this this talk which is I think is relevant, which is like I think it was back at I think it was reinvent. You were we were meeting back at reinvent. So this is
[01:28] like November in in Las Vegas. And you pitched this idea almost this exact idea for this talk even then. And at the time it was like you said there's a lot of unknowns, right? Like people were still trying to figure out how to get any AI adopted across their organizations. Like what you know, how
[01:44] do we unleash these models? The models were changing every week. We hadn't gotten to a lot of the you know, the current state of things which will come back to you. But I I feel like you had a premonition that this is we were headed this way. And so you were I think really kind of ahead of the game, which which is great and I think you put into practice at Workday already a lot
[02:01] of the things that we talked about over the last week. Everything that Ali has talked about around governance and what are the C's? It was like cost and choice and control. This talk is actually a lot about those fundamentals and and how Phoenix you've done at Workday. So, yeah, just like
[02:16] maybe given that, you know, on this slide we've got 40% of agentic AI projects are still sort of slated to fail. What what have you done at Workday? Lessons learned. This is like So, you guys know this is going to be more of a fireside chat, not us just a whole bunch of slides. This is kind of discussion, which hopefully is actually more
[02:32] insightful for you all. So, just yeah, what's what's worked at Workday and what have you learned in your process? What's the journey been like? See, I first and foremost we were not afraid to experiment, right? Like going back to our reinvent conversations like we didn't know what we didn't know. But we had to take some bets, right?
[02:48] Some assumptions, some bets and obviously hedge our bets against that, right? So, things like uh school of thoughts that we form form up on, right? Like one school of thought is like as agents grow, as an organization proliferates through agents, do we need
[03:04] to have an independent data store for all those agents? Or do we need to have singular data stores for those agents? And both are right answers, right? Like I mean yeah, can you create and support an architecture that has an independent data store for every agent or an agentic framework? Maybe yes, of course it's
[03:20] doable. You can also create an architecture that is more unified, more central. But we went back to the fundamentals of data that like, you know, can is it more practical and viable to have a single source of truth for data? And we hedged our bet against that, right? Let's create a single source, a
[03:36] common denominator for our data landscape because it's easier to govern, it's easier to observe, easier to monitor. And and also interoperability became an important part of that decision-making, right? Like as we don't know how the agentic framework, our AI framework,
[03:52] workflows will evolve, but we have to think about a platform that remains interoperable, forward compatible. So, that that was primarily a thought process on how we made these decisions. Yeah, and so just doubling down on on exactly that topic, right? And so, this is sort of a graphic that that you put
[04:07] together thinking about, you know, how do we convert uh the n times m problem, right? If you if you have a bunch of agents and all of them are operating independently in their own data silos. Like, this is not a new problem. Like, you and I have been working in data a long time. The data silos problem is one that we've been, you know, the the whole lakehouse premise was really to start to break
[04:24] down some of those data data silos. We see the problem repeating itself with agents and their agentic context and everybody's sort of having these silos, and I think that you sort of foresaw this and we're working on it with this like, how do we have a shared uh common set of data and governance layer that allows us to actually scale this up. So,
[04:41] yeah, just um sort of thoughts on like, how does that single source of truth work out in your organization? Yeah, it was really like I mean, the mathematicians inside me that that acted pretty much, right? So, do we can we operate in a large enterprise with having a multiplicative problem? Or do
[04:58] we want to convert that multiplicative problem to an additive problem, right? Like, you you know that the picture here tells tells a thousand words, right? If you think about the the the picture with these point-to-point connections, that's a multiplicative problem, right? And you have like hundred and agents and hundred data systems. I'm obviously making up a
[05:14] number. They all connect to each other. That's a very complex spaghetti. Like, as as practitioners, as engineers, how would you manage that complex spaghetti? At one point, you have to untangle that spaghetti in in future, and that untanglement may cost you more
[05:31] more pain than than the taking the pain and doing the right thing at the very beginning. And that led to this idea of like, convert that into an additive problem. And what you see that that that blue uh database sign there, and that's like the common data layer. That's how we
[05:46] call it our universal data layer, Iceberg driven the data layer. That was that converted into an additive problem, pretty much, right? So, now you have an M plus N as opposed to M times N. So, uh that was a deliberate decision and it's it's it's not a new kind of mindset,
[06:03] right? Jason, to your point, right? Like, we have seen as as industry has evolved, as data space has evolved, many, many times the challenges move multiplication to an addition. And so, because this is actually like you've done this in practice, I think the value of this conversation is
[06:20] really, "Okay, we see the slide, we agree with the theory, but the hard part in any organization like how do you do this in practice? How did you actually achieve this or have you achieved this and if so, like how have you done so?" Yeah, I I part of it was primarily driven I've
[06:36] been a lot of guidance came from you, in fact, in in that in that sense, right? Like, let's let's think about what is the most forward compatible, most open data layer, right? What what that that blue blue box would be, right? So, uh we we we started exploring various options and
[06:53] the most viable option for us to create a universal data layer. We named it universal data layer, but what really it is, it's a it's an Iceberg driven layer like Databricks Iceberg layer that that brings that universality to the to the play, right? A few advantages, right? All the data
[07:10] converges into this one central system. Instead of giving data to the agents, we bring agents to the data. And that central layer serve as a universal governance interface, the Unity Catalog, where we could not only
[07:25] do access management, role management, but we can get a holistic visibility of our data landscape. So, that that that was a fundamental shift on how we thought about designing the system. So, few guiding principles, right? Centralization of data, centralization
[07:42] of our governance layer, uh and implementing role management access control through that centralization layer, and make that layer as a as the layer where agents will come and consume data. Yeah, so definitely good
[07:58] good patterns to follow. You talked We're going to like double-click a little bit down into that storage. So, you talked about Iceberg as being the the format you chose to build this universal data layer. Uh and we talked a lot about this as it's very open, uh very forward-looking, right? A rich ecosystem.
[08:14] Uh but yeah, like what are the properties that that you saw around Iceberg that convinced you that was the right uh structure? Yeah, sure. I mean, obviously like asset compliance, like snapshot isolation, and like an open like open and agnostic they're they're they're very important, right? Like I mean, I'm not going to go
[08:30] into the technical details of it. I'm sure most of you would appreciate, right? Like the the capabilities here. But most importantly, right? Like what Iceberg gave us, it gave us it brought oil and water together. Like two opposing teams, they were they converged, right? Now,
[08:46] data data science ML engineering teams, they want to innovate really fast. They want to experiment really fast, and hypothesis testing, hypothesis validation is very very important to them. But if you look now look at like platforms and data engineering team, their work is fairly deterministic. So,
[09:02] how do you merge the world of probabilistic and stochastic world of data science MLE and deterministic worlds of platform and data engineering? And and that's like that's what Iceberg gave us, right? Like it was a format that was
[09:18] equally embraced by two opposing school of thoughts. Like one is oil, one is water, but we made that chemistry work, right? Oil and water mixed, they fused together, right? And that like going like of course like I mean, asset compliance was a contribution to that contributed to that. Snapshot isolation,
[09:35] for example, was a contributor to that, but most importantly like these capabilities was appealing to two opposite spectrums of the polarity who together scale AI in a large enterprise setting. So that that was the the most
[09:50] critical aha moment for us. And you mentioned I want to double click on something. You talked about the sort of probabilistic nature of sort of model driven, you know, uh analysis versus like the very deterministic nature of we're doing financial reporting. The numbers have to be right. Mhm. Uh how how have you uh managed that, you
[10:07] know, that complexity? Still all of this is based on the same source of truth, right? But you have just like different modalities, different personas. How do you all tackle that? And I mean there's no right or wrong answer, uh but in reality we we look at it by separating personas and like really
[10:24] looking at the outcomes, right? So if if we we have to be very deliberate about when we design systems, when we architect systems, when we architect solutions, like what's the end goal? What outcome we are looking for, right? And and bring in the right kind of personas to execute that outcome. If to your point, right? If it's like a
[10:40] financially financially exposed reporting system, then the the kind of engineers, kind of solution architects, they they come to the table, they their uh primary goal is the accuracy, right? Experimentation has less weight to it, right? Hypothesis
[10:56] validation has less weight to it, right? Very very deterministic. You do ABC, the outcome is D, and and it has to be auditable every single time. We should be able to give it to an auditor and they should be able to trace the lineage all the way. But when it comes to like building something completely new and
[11:12] deterministic system, where somewhat margin of error is acceptable, right? And I'll give you a very specific example. Uh what if we are in a situation where we have to uh give guidance to our procurement team, our supply chain teams around
[11:27] um around systems of how they could best negotiate a contract. Right? It doesn't have to be exactly point to point, but it's more of a guiding principle. That's where there is an opportunity to experiment. So, really thinking through like what our risk tolerance levels are and what outcomes are in essential.
[11:45] Based on that, creating teams that operate with different personas that became very important to us. Yeah, I'm going to take that as a a bit of a segue. Um you know, because folks like Workday, people like Phoenix have built these architectures and they really like pushed the limits of what Iceberg is
[12:01] capable of doing, especially for agentic workloads. And it really like identified some gaps uh which is like the next bit of stuff that we're working on uh in the in the storage layer, right? So, all the things that we already talked about are really really important, but there's some things for AI workloads in particular that we're
[12:17] adding down at this foundational storage layer to like make these uh processes better. So, I'll walk through a few of them. The first one is what we call adaptive metadata tree. This is really about unification, right? So, like some of you might be like, "Well, that's great. Workday did this on Iceberg, but I have a bunch of Delta Lake tables. Can
[12:33] I do all the same things on my Delta Lake based tables I can on my Iceberg based tables? Why do I have to think about formats anymore?" We had Ryan Blue was one of my co-founders from Tabular on the keynote with Ali yesterday saying like, "We're not thinking about formats anymore." And here we are talking about Iceberg. But this next version of Iceberg actually unifies Delta and
[12:50] Iceberg so they'll have the exact same metadata structures. So, everything we're talking about going forward is equally applicable regardless of what storage format you're using so you really shouldn't worry about the format. You should worry though about the capabilities that that format unlocks for you and your agents that are working on top of the data. So, this these next
[13:06] two bullet points are really unlocking uh more agentic workflows and workloads. The first one is support for multimodal data types. We've talked to a lot of customers, and Phoenix included being one of those is like, "Hey, I've got a lot of documents inside Workday, right? Workday is full of of documents.
[13:22] And today, my my structured data in tables lives pretty independently from my documents. And that's actually not really how agents want to work, right? We want to have metadata about our documents. We want to build features. We want to do knowledge extraction. We want all of that to live together, both the metadata and the document. We want the
[13:38] governance to be consistent without having to be split across two different things. We want life cycle management to be the same. If I if I remove a record from that table, that should also clean up the metadata document as well and not just like leave something hanging in the system. So, a build ability to add basically documents to our to our
[13:53] tabular data and have it work and operate just like, you know, your integers and your strings do within your tables. Uh that's a big step forward in the next version. And the other one that we see a lot in these ML and agentic workloads is I need to be able to just add columns. This is like a lot of feature engineering or even I'm just again, I'm doing document
[14:09] processing and I'm I'm extracting new pieces of information into a table about that document and I want to be able to add that information without having to restate the entire thing. All right, so unfortunately today with parquet files, you have to add a column uh to your table. That's restating every single parquet file in your table. That's
[14:25] really inefficient. And so, you know, we're building ways that you can just tack on additional features to your tables and leave everything existing in place so we can save a lot of rewrites and get some efficiencies. So, agents really are going to appreciate, I think, the the foundations that we're improving down at this layer. And the good news is
[14:41] like it's not some big change you have to make in order to take advantage of these things. Like if you're if you're Phoenix, you've already got your universal data layer, uh this is an in-place kind of upgrade. No migrations are necessary. These are just be new capabilities that are unlocked uh on your existing
[14:56] which is great. Okay. Uh so, transitioning from that, I want to dig, you know, a little bit deeper into the specifics of of what this agentic systems. We talked about the data layer. We talked about unified governance. Uh but how are the agents really taking advantages? What is What does that architecture look like?
[15:12] Sure. There are few components to it, right? Uh I obviously spoke a lot around like centralized data layer, but it's also important to build like long-term and short-term memories of the agents. In fact, I was in a discussion yesterday and somebody asked me that if I were to really turn the clock back, right, what
[15:28] would I do differently? And what would I do differently is like think about memory management from day one, right? We Again, every bets we make some we those bets are not foolproof and like memory management was something we could have thought more deeply. Right, it it is
[15:43] essential that as various personas they interact through the agent. So, it's not just the data for for the agent to operate, but it is also for the data that capture the personas of the individuals who are interacting with the agents. They're interacting
[15:59] with the AI or the workflows. And that's that's very very critical, right? So, there are there are couple of capabilities that you have in my recommendations you have to you have to think through. One is your long-term memory, right? And that's like you could you could pretty much have it in your
[16:15] iceberg store. And that that's like overall like the persona and the characteristics, the DNA of the AI and the agent you're interacting with. Then there is a short-term memory. Extremely important because that's like the session memory, right? You are I'm giving an example again, right? You log in and you you
[16:31] start just start interacting with an agentic workflow or an agent or an AI and it picks up the context and then it's it it builds the builds its answer or takes action based on that context of the session. And that's the that's a short-term memory and that's that's
[16:46] mostly in in the technical language is the is a vector store. So, something to consider from from day one. And and finally, right? Like the reasoning has to be reproducible and auditable, right? So, there are there are mechanism through which you could
[17:02] actually go and open the hood and start looking at what's the what's the rationale behind the reasoning. And that itself would like the reasoning rationale can become a data store of future. So, that's like if if you know, agents are digital employees like you, uh same like human
[17:19] employees, right? Like when they when they work in an organization, they work with certain characteristics, they leave traces of their communication, they they leave leave traces of their work, and based on that, like their supervisors would analyze and give them feedback and and you know, give you a steer them towards the right path and and steer
[17:34] them towards growth. Similarly, like the reasoning engine is for digital employees on agentic employees, you open the hood and steer them towards the right path and make them even better and better over the course of time. Yeah, and I just I'll highlight here again, like Iceberg does really wear it
[17:51] well as that shared long-term storage memory. Uh I think things like you heard about Omnigen and other products like that allow also uh different agents, you know, operating in the same work stream to also share that short-term memory, so that session-level memory, right? That
[18:06] can be shared not across across human colleagues, but also across agentic colleagues. So, you have humans and agents multiple all coming through this like shared both long and short-term memory space. So, that context window is shared across all parties. I think that's what really unlocks these more complex, you know, things that are actually what we do inside our business.
[18:22] Very very rarely is a single person, you know, working on something end-to-end. So, then we see the same thing in agent workflows. You have multiple different agents are involved, but that context is important all the way through all the way through the process. Yeah, and and and and if going back to our initial discussion on like centralization of the data layer, it's
[18:37] important because if your memory is sitting all over the organization, you cannot really take advantage of it at scale, right? It's like point in time. But in in reality, you want that memory to be centralized, right? And that's that that's what gives you the scale, right? The scale doesn't come from okay,
[18:52] you have you have used the most most frontier models and it's the most advanced model. Scale comes from how you design your architecture. Yeah, if you have all those uh interactions with your AI and with your agents, and yeah, they they're locked up locally somewhere or they're very spread out, it reminds me of like, you know, the spreadsheet days. Everyone's got
[19:08] spreadsheets and they're all over the place and no one knows where to find all the data. It's like very disparate, right? Exactly. We bring it all together into like a database or a or a lakehouse. Similarly, with like your agentic memory and context and and everything that goes along with that needs to come together so that you can keep track of it and and leverage it. That that becomes a feed a
[19:24] positive feedback loop. Exactly. And it's the same reason why like organizational leaders, right? They want to break human silos. Teams should not operate in silos. If teams should not operate in silos, agents should not operate in silos, either. 100%. All right. Just maybe for you to dig in
[19:39] a little more here about, okay, that's cool. That gets you this foundation, a very strong foundation. And then the next, you know, phase of it is like scaling this thing up. So, you know, again, great in theory, but how does this thing actually work? How does it scale for you? What what new things did you uncover as you try to really scale
[19:56] this thing out inside Workday? The many things, right? Like that and and and the picture there, right? Again, telling a thousand words there, right? It's Yeah, it's you can you want to experiment with one agent, right? Like connect some data sources to an agent. Great. It's it's easy. You can very quickly build something, right? In
[20:12] today's day and age, right? You want to build an agent, put some data to it. You can you can do a fairly good job, right? And and probably like while you are listening to this conversation and you can probably code it like build something on your cell phone, right? It's that as easy as that.
[20:27] Then you go to a 10 10 agent scenario, like, you know, 10 point-to-point connections, like, you know, multiple agents. They are talk sharing some common data stores, sharing some common repositories, and they are fairly, you know, fairly easy way that talking to each other. You can manually keep track of their records and their
[20:43] registry. Um good enough, right? You are you are probably operating within a three team of like 20 people. They are doing some work. But then, in reality, uh because building software is easy, relatively speaking, building agentic
[20:59] frameworks are relatively easy. You know, AI has definitely empowered and democratized it. What about 100 agents? Like now, look at look at that diagram. Like it it that is a reality and we realized it that we cannot operate like this, right? Because
[21:14] it it will be again going back to my spaghetti example, it will become a spaghetti and it will be impossible impossible to really operate with confidence and confidence and uh governance. Um in theory, right? If you if you have
[21:30] that 100 agent scenario, point-to-point connections, it might work technically. But someday some agentic workflow, someday some AI will give you a wrong information or make a wrong decision and the whole trust
[21:46] plane around AI will come down the drain, right? Like 100 good things but one bad thing and it will if the the trust trust erodes. So, you don't want to be in that in that situation where like the organization trust starts eroding on your agentic framework, your agentic ecosystem. And that's why like
[22:04] the what you see at the bottom is is is very very important, right? Like you go from that spaghetti to you go to a governed phase where all the agents like bringing agents to the data, right? You have this data layer, agents come to it.
[22:19] So, in a way you pilot with one agent, you get into some frictional phase with 10 agents like you know they are interacting with each other like the things are heating up but really the scale start from 100 agents and once you are like in that territory like it the obviously these numbers are
[22:35] arbitrary to kind of convey the concept but but once you are in the scale territory, the design architecture of point-to-point connections will not work and we we learned it the hard way and and we redesigned our thought process around it. It is working like a magic
[22:52] for us. Obviously, there is lot work lot more to do. We don't know many things still but the whole idea is like take calculated risks and if something were to not work like fail fast fail forward. Yeah, I think that the trust is really really important. You mentioned it. It's
[23:08] difficult to build, easy to lose, especially we're all very skeptical. Exactly. Like even in the early days of BI, right? Like you could generate like hundreds of dashboards, but one dashboard will give that wrong piece of information to someone sitting in in a C-suite of the organization, they will trust distrust
[23:26] all the dashboards that come into their desk. And that will that might happen in the agent space, too. That's actually how we got Iceberg. That's a different story. I'll tell you sometime, but Okay. Okay. I I I love to hear that. a it was a a membership report from Netflix that went to Reed Hastings that was wrong. And then then we got then we
[23:42] had hey, we need asset transactions on our lake. Anyways, uh different story for another time. Okay. And so we mentioned all of this obviously, you know, scale, governance is a really more important part. And when we say governance, yeah, it's it's not just the access to data. There's all
[23:59] the other pieces that you've been talking about like how do you monitor, we know things are eventually going to go wrong. How do we have how do can we can we run that down and make corrections? So what have you all done in this space to like really put a bow on this whole architecture? Yeah, I I guess like it's not just like centralization of the data bringing
[24:15] agents to the data. It's not about like solving the multiplicative problem to an additive problem. It is all around also about like making sure that we understand the data landscape that we have, right? And making sure that there are it is compliant and making sure it is it
[24:31] doesn't lead to leakage. And that that's where like Unity Catalog came as a as a rescue for us, right? Pretty much. And and and and a few things here, right? Like one the kind of data we deal with like some of them are sensitive. Some of them
[24:46] might be sensitive from a perspective of PII. Some of them might be sensitive from a perspective of socks and ITGC and whatever and whatnot. So understanding like what tags we are operating with the data, tagging the data, right? When you go in and see your data landscape, you
[25:01] should be able to understand very well like what is data can and cannot be used for. So, that's one dimension of it. And then there is a dimension of uh that's all the that's the compliance dimension. Then there is a dimension of like, you know, lineage and auditability of the of the data. So, at the end of the day, right? Like if if
[25:18] that mistrust scenario happens, and if somebody tells you like, you know, your agent is lying, you should be able to trace back and say it. Yes, the agent is lying and why? Or the agent did not lie, hence this is the proof why the agent did not lie. And that's very very
[25:33] critical, right? You should be able to like, you know, trace back the data through which the agent is agent of the AI is giving you the outcome. Then there is a important component of security, right? I'll I'll give you a very specific example that that will apply to almost every industry, right?
[25:49] So, almost every organization has a sales team. And as you're you're thinking about sales, right? You you have territories, right? Let's assume like an account executive who works in the north territory, and there is another account executive
[26:04] who works in the south territory. The north territory account executive shouldn't be able to ask the agent that tell me the data tell me the account status, the health of the south territory. They should agent interface should know the entitlement of that individual and block
[26:19] the information that they are not entitled and authorized to see. It's it's it's no different from from the fact that when you log into the system, the system actually impersonates and understands who you are, what your role is. Based on that, it allows or
[26:36] disallows certain accesses to you. So, the same applies when you are interacting with an agent, right? Like you shouldn't be able to prompt inject it and get the information that you are not authorized to see. That that that will be a debacle, right? And especially like when we are operating
[26:53] with highly sensitive data. What is human What about human resource data? Financially sensitive data, market data, and whatever and whatnot. And that's why it is quintessential to think that data security, role management as you are designing your system. It's not about
[27:08] really like, "Okay, take some data, build bring an AI, build an agent, build an a workflow." No, it's it's also about thinking more holistically. And Unity Catalog was a key contributor to achieve true enterprise scale for us when it comes to
[27:24] compliance, security, observability, lineage tracking, and monitoring. Yeah, I would love to be able to take credit for like all the innovation that happens at Databricks, but the reality is, you know, a lot of what you saw this week, a lot of the announcements we made are you'll see them like represented here. And really it's like partnering
[27:41] with with people like Phoenix. Like they're kind of running ahead. They're hitting these problems ahead of time. They're solving them in their organizations. We're learning as they go from all of our customers who are hitting these problems. And then we're really just trying to bring those learnings back and and bake them into Databricks products so that, you know, everybody can can share
[27:58] in the experiences. So, thank you for being a great design partner. And and to that end, like what what comes next, right? Like where do we go from here? Yeah, of course. I mean, these are some some directional examples. Like just just to be clear, like they we obviously obfuscated the real numbers for
[28:14] confidentiality reason, but really this is the art of the possible, right? Like you could you know, you you could build a bot, right? Now that you Now Now that you have your data centralized, your governance centralized, your agentic interface can really scale. That These are some of the
[28:29] outcomes that you can potentially achieve, right? You could accelerate your finance closeouts or recycle. You could like look at look at your attrition rates way ahead of time. You can You can improve the way you do supply chain management. Like one of the complexities of large complex enterprises are like their
[28:45] procurement team, their supply chain team has to look at so many contracts across the across the board and sometimes even at a global scale, right? So, ensuring that you could you could get best out of your negotiations, that's that's possible. Then, obviously marketing and sales use
[29:00] cases where your retention and your sales planning can improve improve a lot. So, there's So, among many, these are these are the art of the possible and directionally we also applied our our capabilities towards these functional areas.
[29:17] Uh and really like Jason, to answer your point, right? What what what next? I think we are just getting started, right? This is like the iceberg was really like at the like like just we just what was a drop in the ocean. There is so much more now that we have to think,
[29:32] right? Like agent-to-agent communications and like you think about like protocols and and MCP servers, registries. Um so, yeah. Yeah, there is there is a lot to be done, right? If we had to really scale, if 10 is the best state, we are probably
[29:48] like at a three. Okay. And that's what's exciting. That's that's what makes it exciting, right? Like we have to cover like seven more steps. So, if I if I happen to be not even on step three yet, I'm going to say I'm still on step zero or step one. Like what would what would be the Point taken. Yeah, yeah. What would be your recommendation in terms of like what is
[30:04] the what is the next thing to do? Like you said this is like very much the art of the possible, uh but I mean this took you guys, you know, a year plus to to to to build out. But like where do I start? Where do I start? I think from a next steps perspective, I would start looking at
[30:20] three things. First, I'll start looking at um more towards interoperability. You touched some of it, right? So, iceberg was a great great starting point. But organizational data will be like lot more complex. It will be coming from lot
[30:36] more sources. And uh and uh it is just it's not if, it is it is when we these agentic interfaces would have to interact with data and systems that are very wide, right? And and that leads to the second point is like establishing
[30:52] like registry of agents, establishing communication protocols between various agentic interfaces would would be would be the next obvious choice. And then there is an obvious choice around uh governance, observability. I mean some of it is even announced in the the kind
[31:08] of conference like governance, observability, automation component. Because it's not like okay, you build things from ground up, right? You you have built now, but how do you monitor, observe, automate those components? So, the you know, what you see here is like really how systems will we believe the
[31:24] systems will evolve over the course of time. Like there it's not just agentic workflows, it's applications and and more you know, model communications and uh external systems. They all have to coherently operate with each other. And and more often than so, we will
[31:40] experience that uh systems that are vertically isolated will have to talk to each other. Like it's it's going back to the you know, early days of software engineering. Like there was like you know, two softwares uh provisioned by different organizations, they will not talk to
[31:56] each other. You cannot communicate. And then ideas of APIs came into play. But I guess like that the history is kind of repeating itself where these agentic interfaces are available and proliferating everywhere. But for them to truly work coherently, they need to
[32:11] start talking to each other. So, we have to not just think as like going back to my example of like, you know, deterministic and probabilistic. So, while agentic activities themselves might be determinist it might be probabilistic, but the interaction of those agents, the UX UI have to be
[32:28] deterministic. So, they're bringing both the worlds together. That completes the architecture. And that's why I I firmly believe that it we we just getting started and we don't know a lot. So, like ability to try, test, fail, re- retry, and really create the framework
[32:44] and the blueprint of what makes it successful is important. Yeah, and I I feel like it's a theme that I've heard throughout the week and certainly I heard earlier today in this talk as well, which is um kind of just it's worth it to invest in these foundations, right? Like it's easy to say, "Oh, I just have this one use case. I just want to connect these
[33:00] two systems." Like you say, "Oh, point-to-point connection. Oh, they both have APIs or MCPs. I'll do these things communicate with each other." But that quickly becomes two, quickly becomes 10, and now all of a sudden the spaghetti starts to build up. And so, to avoid that, you know, invest in building the foundations, bring all the data
[33:16] together, right? Put it in open format, invest in a central governance layer that all that communication flows through. You can still connect all these things, right? And then once you have that foundation, you can start making these connections much faster, you can start to experiment much faster, but you you know you're not creating a big big mess that tech debt or something you're
[33:32] going to have to clean up. And also like when you do need to swap out a model, put in cost controls, you'll you'll have a natural choke point to put that. If you don't If you don't have that in place, then it's really hard to come back and retrofit it later. Exactly. So, I I I'll add two more points. Like I I really like the way you called out swap swap out a model. So, we
[33:48] have to be very deliberate to design our systems like an architectural modularly, right? Like we are we have to think like we are builders of a LEGO toy, right? You take the parts and build build a toy. But you equally if a part breaks down in future, you should be able to swap it very quickly, right? So, that's
[34:04] very important because it's still an uncharted territory. And the another analogy I will draw, right? Like if you were to really start a company, and it's a company of 10 people, they will probably communicate well and you will be quite quite nimble and able to move things fast. As it becomes a 100
[34:19] employee company, there are maybe two, three, four departments in the company, there is some friction you would experience like there might be some politics you will also experience, right? People liking, disliking, every people emotions. But the moment it becomes like a thousand employee company, 10,000 employee company, lot many
[34:37] frameworks have to bring to the you have to bring to the table, right? How maybe offsites and maybe how your employees talk to each other, policies and practices. It's no different from for from the future of agentic world, right? You We are small right now. Many many
[34:52] organizations, many of us are starting small, but we need to be quite thoughtful and deliberate about the fact it will become a large enterprise, like him same taking taking that like startup example that your startup will become an enterprise. And when it becomes an enterprise, you cannot solve the problem
[35:08] then. You have to solve the problem through the steps. Yeah. Yeah, unlike an enterprise, like hiring actually happens relatively slowly. Like your agents hire will hire very very quickly, right? The hiring cycle for an agent is much lower than a human. So, these things will proliferate quickly. Okay, thanks everyone. We appreciate the
[35:23] time. We'll We'll hang around for this question. Just come to the front. Thank you, everybody. Woo!

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.