Sponsored by Atlan | The Enterprise Context Layer Why Agents Fail, and the Fix, Demystified & Demoed

Summary

Gartner declared 2026 the year of context, and Atlan's demonstration shows that the gap between a working AI pilot and production is a context problem — agents cannot answer reliably when they do not know which revenue definition to trust or which tables are authoritative.
The enterprise context layer sits on top of Unity Catalog and comprises three layers — AI-ready data with automated lineage and definitions, semantics and ontology, and skills — built by reverse-constructing the institutional knowledge data teams already own.
A live demo shows dark Databricks tables enriched by Atlan context agents with automated lineage, AI-bootstrapped definitions, and a Context Quality Score, resulting in a Genie query returning a governed response via MCP and moving AI autonomy from 20% to over 90%.

Chapters

FAQs

What is an enterprise context layer and why do AI agents need one?

An enterprise context layer is a structured repository of institutional knowledge about data — including lineage, definitions, business rules, and semantic metadata — that gives AI agents a governed understanding of what data means and which sources to trust. Without it, agents cannot reliably answer business questions because they do not know which revenue definition to use, which tables are authoritative, or what relationships between entities mean.

Why do AI pilots fail to make it to production?

This video cites Gartner research showing that only one in five AI use cases makes it to production, and 56% of CEOs report zero financial benefit from AI. The core issue is not model intelligence — models are improving rapidly — but context readiness: agents cannot perform reliably without the deep, governed understanding of business data that human analysts carry implicitly.

What does the Atlan and Databricks integration demo show?

The live demo shows Atlan context agents automatically mapping lineage, bootstrapping business definitions, and computing a Context Quality Score for dark Databricks tables that were previously unusable by AI. After enrichment, a Genie query against the same data returns a governed, accurate response through an MCP connection — demonstrating the before-and-after impact of context enrichment.

What is the Maya framework and how does it explain the human-vs-AI performance gap?

The Maya framework in this video describes what makes a great human analyst effective: not raw IQ but the ability to understand the business, learn from feedback, and deliver outcomes in the real world. The point is that less than 10% of job performance variance is explained by cognitive intelligence — performance comes from contextual knowledge and adaptability, which is exactly what AI agents lack when they have no enterprise context layer.

Full transcript

[00:09] So, we're here to talk about context. How many's heard how many times have you heard the word context today? One, two, three, 10, 100. I believe if there was a context drinking game at this conference, um, we
[00:24] would all be pretty wasted. Uh so actually exactly a year ago at uh summit uh we had this slide on at our at our
[00:36] keynote and we said at the dawn of the internet era Bill Gates wrote this very famous blog uh saying content is king in the internet era. Um and we we said in
[00:50] the AI era context would be kink. Um it feels very surreal that that was less than a year ago. Uh since then, March of
[01:02] this year, Gartner went up on stage and said context is king. 2026 is very much the year of context. Um right from context graphs to it being the next trillion dollar opportunity to what does
[01:14] the company brain and the AI operating system for companies look like. Why is context in the spotlight all of a sudden? Um I live here in San Francisco and I feel like I live in a constant
[01:28] state of confusion. Uh because on one side I am around these people who are talking about AGI and they say you know like AGI is going to be here in you know a year. Um and on the other hand I work
[01:41] in the real world and then I ask my customers and I say hey how are things going with AI? And typically their eyes will droop down and they'd say something like, uh, there's one P that made it to
[01:54] production. Uh, you know, they definitely say something about security. Um, sometimes they'd say, well, it's a lot of hype.
[02:08] On one hand, and there's no doubting this, the models are getting exponentially smarter. Two years ago, they couldn't pass the bar. Today, top 1% of test scorers. but at the same time not exponentially more useful. 56% of
[02:20] CEOs say zero financial benefit from AI. Uh only one in five actually make it to production.
[02:33] back to humans and try and understand what makes humans work. And hiding in plain sight is somewhat of the answer. Cognitive intelligence or IQ doesn't
[02:46] really determine real world effectiveness. Less than 10% of job performance variance is explained by IQ. Think about it like would you say that your best teammate uh your best employee
[02:58] is the person who also scored the highest on the SATs or would you say you know this is the person who learns the most and understands our business and they take feedback really well and they work really hard
[03:11] in the real world we care about performance performance are outcomes that you deliver in the real world and performance is a function of two things on one hand it's a function of intelligence cognitive of horsepower.
[03:25] But on the other hand, it is a function of context, knowledge, skills, and tools that are built doing the work. This is what they they say learning on the job is about.
[03:39] we have compounded in one of those dimensions. We have thousandxed in intelligence. On the other hand, context barely moved. We've moved some data to the cloud. uh
[03:52] but outside of that you know the knowledge of your business it's the logged in dashboards and slack threads and the analyst who might be leaving next week this is what we believe is the next frontier contextual intelligence
[04:21] intelligence systems but higher context systems, right? We had um you know well tuned machine learning systems actually relatively high value. You move
[04:33] to the other side with high high intelligence systems with low context you're actually in danger and so in an ideal case what you want to get to is high intelligence and high context systems.
[04:46] So, normally when I when I tell people this, they say, "Okay, well, but like what does this actually mean? How do we help AI build context about our businesses?" Again, I always like to go back to humans. Um,
[04:59] I want to take you into the life of Maya. Maya is a data analyst at let's say Context Burgers. Uh, and she's the one analyst that every person pings when something is off. Uh, everyone knows
[05:13] that analyst in in their companies. Um, this morning it's Marcus. It's a franchisee with, you know, 14 stores. And the question is, why is my drive-through time up this week?
[05:26] Sounds like a pretty simple question. It's actually a pretty hard question to answer because as Maya starts trying to answer this, she needs a bunch of context. First, she needs to understand knowledge. Then there's expertise, and
[05:40] then there's norms. knowledge. For example, she needs to know well what does drive-through time actually mean? She also needs to know what does this week mean? Monday to Sunday like you know um or Monday to Friday.
[05:57] But then you know the great analyst they don't just answer the what question they answer the why question right uh they understand seasonality they know that in the first week of you know the third quarter we typically typically have a seasonality event. They rule out that
[06:10] premise and then you know she there's rules of acceptable action you know give Marcus only data for his 14 stores don't give him data about the entire like business knowledge expertise and norms these are
[06:23] the three kinds of things that a great analyst needs to answer the why questions Maya nails it you know she says uh your drive-through time is up 18 seconds but only at single lane stores um she finds
[06:36] the reason there's a new burger that we launched on Monday. It takes extra prep. She gives the answer to him. She gives him a fix.
[06:48] But how did Maya learn to be a great analyst? Sure, Maya joined and uh she had some data dictionaries and she had, you know, some documentation that she started with. Uh but then Maya, you know, went
[07:01] and you know, started shadowing some people in her team. She started doing some analysis work. There was a senior analyst who pointed out to her that hey we have a seasonal business in this quarter. Uh and then she made some
[07:14] mistakes. She learned some wrong answers while she made these mistakes. She realized never to do them again. Um her manager gave her feedback. Uh this is how a great human learns on the job. And
[07:28] so the question ahead of all of us is how can we go build Maya
[07:40] such that like AI AI today is what Maya was on day one right uh how can you onboard Maya Jake Diego all these people in your teams um as you start building an AI native firm that is what an
[07:52] enterprise context lure is meant to solve at the very simple like player. It is a system that turns knowledge, expertise and norms into machine usable context
[08:04] for AI. Uh the way I like to think about it is it is a organization's living brain.
[08:17] Typically when I when I say this mostly you know people say okay well palpa I am bought in I'm technical. Can we just go into like the details like what is an enterprise context layer? So, we're not going to tell you. We're going to try and show you. So, I'm going to bring up Tom and we're going to through today
[08:30] actually try and show you the different parts of what it takes to build up enterprise context inside organizations. >> All right, let's do it. >> It's funny the distribution of people.
[08:42] Like, it's really heavy over here and it's gotten really light over here. So, if you need a seat, there's plenty over here. The ship is definitely capsizing one way. Awesome. All right. Um, let me switch over so you can actually see
[08:55] what's on my screen. Now, I know it might be a little far for some of y'all to see. I do this on purpose so you can stop by our booth afterwards and have a really good demo that's personalized for you. But what am I going to talk about today? Um, we're going to do this sort
[09:08] of in tandem. I'll show you a couple things. We'll bring Pulpa back on to explain a little bit more. But where we're going to start first is sort of that end product. what are we trying to achieve and what is knowledge and
[09:20] expertise and um and our norms along the way. So on the left hand side this is sort of an agent comparison. On the left hand side as you start in a company and you start building agents the first
[09:32] thing that you think of is I need a fantastic prompt. Everybody knows about prompts. I feel like that's so 2023, right? You build a fantastic prompt and then you say it needs access to the data so I should probably give it the schema.
[09:45] And that's where most companies start. You give it a schema. You give it a fantastic prompt. On the right hand side, we have it hooked up to the ATLAN context repository. And the context repository consists of that knowledge,
[09:59] that expertise, and those norms. And that takes the form of personas, markdown files full of skills. Semantic models are a piece of the context layer. And what I'll walk you through is on the
[10:12] left hand side that agent that has access to the schema that has a fantastic prompt. It does exactly what you would expect it to do. It goes and starts querying tables and then gives you a response that you would expect.
[10:24] It's actually an extremely familiar response to me where I ask it a question. I realize it's a little vague and so does it. And it responds back with I don't really know what this means. I don't know. In this question, in this case, I asked the question, why
[10:37] is my drive-through time up this week? Same question that we saw my answer here. And it asks, well, should I be using this column? How should I be thinking about calculating this? If you could provide me a little more context, I'm happy to get you an answer. That's
[10:50] how most AI responds. So, I'll go back and forth and I provide my personal context to this and I get an answer. But personal context doesn't work when you work in an organization of 90,000 people like Merly or other companies out there.
[11:04] I call you out. See, I knew I'd bring you back up in my presentation. Uh, but on the right hand side, what happens? The agent first goes out and gets all of that knowledge, gets the expertise of
[11:16] the best analyst in your business, and it realizes that if I'm going to do this drive-thru time analysis, I need to first look at things like seasonality. I need to perform a whatif analysis. I
[11:28] need to understand what these terms mean in the business when they mention, for example, operators. What is an operator? Is it the person that is cooking the burger? Is it the person that owns the store? Are they known as operators?
[11:40] These are all things that are missing from publicly trained LLMs. After it gathers all that, it queries the information that's needed. And it gives you not just the answer of the reason your drive-through time is higher
[11:53] is that you introduced a new burger. And that new burger takes a little longer to to make, but it also goes through and realizes that there's different types of stores. You have a single lane store versus a dual lane store in drive-throughs. And the dual lane stores
[12:05] are totally fine. And the icing on the cake, especially in the data world where we're trying to tie AI use to actual company benefit, we're saying you're going to lose 11 to$14,000 a week if you don't solve this problem in your
[12:18] organization. So, as you're working through use cases where in the data world where you're trying to bring value in AI, these are the things that you need to think of along the way. Now, I'll flash up quickly here a context
[12:30] repository to show you what this actually looks like. And then, of course, we'll walk through how you get there in just a little bit. But the important thing to take away here is that it's full of different types of context. It could be semantic models.
[12:43] So, please don't come up to me at the booth and say, "What's the difference between you and a semantic model?" It's a piece of the context layer. It gives you some of the answer. There is a lot more along the way after this. You should understand it. And the most
[12:56] important thing about your context repositories is that they need to be open and portable. You will have tens of AI building platforms in your organization. I'm sure like I talk to
[13:08] people all the time that have writer, they have Sierra, they have agent bricks, they have all these different tools that are thirsty for context. And so they build context in these tools individually and don't manage it at the enterprise level. And then when you
[13:21] switch, you have to pull it out and figure out how do I pull it out first and put it somewhere else. At we believe the standard is keep it open, keep it portable. We store it iceberg native on top of S3 so it's available to you
[13:33] wherever you need it. And it's oneclick deploy oneclick deploy to things like datab bricks genie spaces out to snowflake to things like claude to open aai with your chat GBT through MCP as
[13:46] well as uh your Google ecosystems and anything that accepts MCP at that point. There are so many different ways of communicating with the context repository. Our goal is to keep it open and portable. And with that I will turn
[13:58] it back over to PK. >> All right. So, normally when you go and see one of those demos, every technical person I've spoken to
[14:10] says, "Yeah, it works in a demo and not really in real life." Um, over the last year, we've been working super closely with customers on what it takes behind
[14:22] the scenes to actually build the context layer that it takes to productionalize AI.
[14:35] have hundreds of agents and dozens of agent platforms. At the same time, I have a heterogenous mess internally of context. I have systems of record. I have systems of knowledge. I have systems of data. I have systems of of BI. I have all these
[14:49] systems on the left.
[15:02] get to. Context is often made of three things. The first is what we call AI ready data and the knowledge graph. This is really where
[15:14] an agent needs to understand what your data means. You know, u how do I think about how do I filter on this on this data set? How do I join on this data set? What are the most popular business questions that are associated with this
[15:26] data set? What does this column name mean?
[15:38] knowledge in your business. The second layer of this is what we call semantics and ontology. There was ontology that that Ali talked about today in in the keynote, right? What does a metric mean? How do we define
[15:51] this metric? How are these different metrics interlin one to another? uh how does finance look at drive-through time differently than my supply chain team? Uh and then there's ontology. This is really how like my business is connected
[16:04] one against the other. Uh how do I understand the relationship between product and store um and individual manager and all these things that I understand about how my business works.
[16:17] The third reproducible unit is starting to be skills. Uh this is really where expertise and norms are starting to get encoded. It's an open reusable format um launched by anthropic uh but now has
[16:31] been adopted across the entire ecosystem. Think of it as a way to take process and norms and procedures and convert it into a reusable unit uh of of knowledge inside your organization.
[16:47] In the journey to build this context layer, our customers often go through three kinds of problems. In the early days, context bootstrapping. Today, it's very easy to build an agent. It's really hard to give it the business context. The thing about uh AI that's different
[17:01] than BI is that in BI, you knew what questions people were going to ask. It was bounded. The minute you open it up to an AI ecosystem, your end users are asking all kinds of questions. There is
[17:13] no end state. This is also a non-deterministic system. I've had customers that have been stuck for over five months in this testing hell where someone like they launch something and then somebody asks one question that it
[17:26] didn't give the accurate answers to and trust breaks just like that.
[17:39] business changes, your systems change. Uh when something goes wrong, well, how do you even trace back what went wrong? Was it the model that changed? Was it my eval that changed? Was it the harness that changed? Was it my context that
[17:52] changed? And the third state um as you start going across this journey we start seeing customers go through the same problem that actually in the data world we used to have when we used to say uh
[18:04] when I look ask for the revenue number sales and finance always give me two different numbers of dashboard sprawl. We're seeing the same problem with agent sprawl uh where you're deploying all these different agentic systems. The
[18:16] worst part about these systems is each of these have memory so they also learn differently. And so I you know we've had customers who've said things like now one of these agents learn it doesn't
[18:28] pass back to the others uh and now all of them speak a different version of the truth and finally the world is changing so fast uh in our company u started this
[18:40] year every engineer used cursor um February it was almost all clot code and now we are 50% clot codecs uh the world is changing so fast that
[18:53] you're constantly also starting to shift in the AI layer. In all of this, we're hearing customers say context is my IP. It's the one thing that is is core to my business. I don't want to start getting logged up into any of these individual
[19:06] architectural patterns because change is the only reality in this world.
[19:18] The first there's a lot of business context that exists today in systems. Systems of record, systems of data, systems of knowledge. The worst part is every time data moved, context was lost.
[19:30] Right? Like if you take context from your system of record to your data warehouse to your BI tool, every one of these steps context was lost. One of the things that we've realized is if you're able to reverse construct these systems,
[19:42] you can actually go back and reverse construct a large part of your core business. The second the context development life cycle. Today we have this once in
[19:54] a-lifetime opportunity where we are building agents and as a part of this agent development life cycle there is a context development life cycle. This is an opportunity to take human tet knowledge that was always a challenge
[20:07] and codify it.
[20:19] that part of the agent development life cycle. The third compounding learning loops. The beautiful thing about agentic systems is that every time a user interacts with it, it creates context
[20:34] and you can take that and use it to create compounding knowledge systems. For example, with a dashboard, you would never know if a end user got an answer or not. Now, if you would roll out one
[20:46] of these agents and they say, "I asked for revenue by the enterprise segment, not by, you know, the mid-market. Can you remove it?" Now, you know, and now you can say, "Hey, the next time this
[20:58] user asks for revenue, they typically mean revenue by enterprise. this user is in my enterprise team in this region. Maybe for their entire team, the next
[21:11] time they ask the question, I will probably give them revenue by enterprise segment, right? You can start building these context loops over time uh that compound over time. Finally, context activation. How do you
[21:25] open this across all of your ecosystems? And underneath all of this context quality, context drift, a lot of the problems that we've seen in the data world are starting to get recreated in
[21:38] the context world.
[21:52] construct from your data layer, we've been able to see that you can go back and create back a ton of um of your business. I'm going to start by actually playing a video of our customers as they
[22:04] experience this and then tell you how we went about doing it. >> Oh, think we need a little music back there.
[22:18] >> Uh is there music or sound? >> Got it. Go ahead and play. >> I don't want to have to sing it, guys. I don't have a good voice. So, >> no.
[22:34] >> Okay, >> you can just Why don't I show it live, PK? >> Yeah, >> I'll just show the context agents when we get there. Yeah. So at a high level the reality is if you and I throw some tables into cloud code today we will get
[22:49] something that looks like an ontology. The challenge is it's not accurate. So the question really is what does it take to create something that's accurate? We've been at this problem for
[23:02] many years. We launched Atlin AI in 2023. Um, and I think we've eventually gotten to the other side of the tipping point of what it takes to make this accurate. In this journey, we have realized context quality really compounds. So,
[23:16] the way we think about this, let me take something as simple as a column description, like writing an accurate two sentence column description. If you're able to understand lineage and say, okay, where does this column come
[23:28] from? All the way from the system of record and all the way down to a BI tool. Then you're able to say, "Hey, who's using this in my BI? What are they using it for? What does that business do?" If you're able to read through all
[23:40] your SQL queries, you're able to today reverse construct a column description that is better than what a human analyst would write. 89% of our customers say that this is better than what a human would write. In fact, I'm having customers now ask us to um essentially
[23:55] um overwrite what the humans have written with what the AI has written, right? Um, then if you're able to take a really high quality column description and connect it to BI semantic logic, you're
[24:08] able to start reverse constructing things like filters, top business questions. Why does someone use this? You're able to take those together, you start bounding, you're able to create really strong domain bounds. You're able to take that, you can start reverse
[24:21] constructing metrics and KPIs. You take all of this and then you build a great harness around it. Then you're able to start saying okay this is how my business works. This is entities. These are relationships. It's the holy grail
[24:33] of ontology. That's the journey that it takes to compound context quality at every stage of the pipeline to get highquality context where you can actually roll out what the AI does in a way that is high
[24:46] like that is accurate and trustworthy. So maybe Tom, why don't you actually show us what this looks like in action? >> Yeah, let's do it. Who doesn't love live demos? See how this works out.
[25:01] All right. Um, well, let's start at the beginning. We talked about the data layer or we talked about building the foundation and context mining along the way. At Atlin, we have hundreds of connectors. They're as simple as enter your credentials, set a schedule, click
[25:15] go, and it's going to constantly bring in new metadata. And think metadata or context like your object tags, your descriptions. If you've invested in writing descriptions inside datab bricks
[25:27] already in Unity catalog, it's going to pull in all the SQL query history and we're going to use that along the way when we build context. And what you end up with is a little bit of what I like to call your enterprise data graph. So
[25:40] this goes from not only your data warehouse but outside that data warehouse to systems like we see Netswuite down here or if I move over to the left your systems of record like your SAPs.
[25:52] And as I scroll over to the right, we even have that BI layer that Pulpa was talking about. And I can continue to get all of that rich context that lives in the BI layer. Let's face it, we've spent so much time building dashboards to
[26:05] answer business questions already. It's such important context to bring in to start to do that context mining and build that solid foundation of the data layer. And of course, this isn't just at
[26:17] the table level. I can go down to a column level and see exactly how columns move throughout the system. This is what AI needs to be able to look at a table. It discovers, finds subscription health for a customer, and then knows how to go
[26:30] upstream or downstream depending on if it needs more information. But this is just the data layer. We're going to get to the business layer here in a sec. Now, from here, the largest problem that
[26:42] we run into is the video that we were going to show you, but it didn't quite work for the sound. This is what we're going to do live. What we do is we take this information and we help you
[26:54] document your data estate. This is what happens when customers start to connect. First, you connect, you crawl in your assets, you get super excited, but you have blank descriptions in your columns, you have blank tables,
[27:06] there's no rich metadata along with it to understand how the business is using this. And what do you do? We've been a company for six, seven years now, and we've gone out, we're top right in every quadrant that you could possibly find,
[27:20] but we're still having to find ways to gamify and get humans to go in and fill out this context. It's the biggest struggle in an organization. especially if you're coming from a large enterprise. Well, this is where we introduced our context agents and
[27:34] context agents take that foundation or that um the compounding nature of that context to build in specific uh types of context. So, if I just go into one of
[27:46] these buckets here, I'm taking a look at some data bricks assets that I want to enrich. I'm going to enrich things like descriptions, like a readme, and even SQL intelligence to build a beautifully
[27:59] enriched asset that AI can then use to answer a lot of my questions. And if I go to an asset profile or a table, in this case, a data bricks table, and I take a look at the description again,
[28:11] I've already run this, and I have a beautiful description, great column level descriptions, and I didn't stop there. I have a fantastic readme that tells the bi tells the agents exactly how the business uses this table along
[28:24] the way. For example, what are the popular business questions that are being asked and one of those categories was our SQL intelligence. This is where we actually go through the query history inside your data bricks inside whatever
[28:37] ecosystem you have SQL server and we reverse construct the business questions that were being asked of this table. We look for things like relationships between tables, filters that are common.
[28:50] We all know if you're doing some sort of analysis, there's always which date filter do I use along the way. We source that information through all this compounding metadata that we have. And I've just talked about the data layer.
[29:02] So at this point, we have a data layer. We have all your uh enterprise data graph crawled in. We know how the data flows through your system. What are we actually doing for the business at this side? because there's a whole bunch of
[29:14] logic out there on the business side that's going to be important to tie back to this data layer. Well, it turns out these context agents can also be used in a compounding nature to build an ontology where we use your data graph
[29:27] full of the assets that you have and all of the knowledge files that you can upload. A knowledge file for us is think of like the gold layer for your unstructured data. things like SOPs, how your company runs, metric calculation
[29:40] rules, and we run those agents on top of that beautifully enriched data layer to build an ontology. And in this case, I'm looking at a specific entity to show you
[29:52] what this looks like from a business standpoint where we have a subscription entity. It has these different properties. It is buil by the invoice entity and it even has things like six
[30:04] different metrics that the business has agreed upon. This is exactly what an agent needs to be able to understand how these entities in your business relate to one another. And really the cherry on
[30:16] top to all this is we take that logical layer and we link it to the physical layer to show that agent where it can go query for this for example here these data bricks tables to find that
[30:29] subscription entity and run the the query that it needs on top of it. Now that's the last little part here as I look over at time to turn it back over to you Pulpa and I'll come back one more
[30:41] time. So now let's go back to our analytics AI agent. So far all this work we've done, we've done only one part of the three circles which was knowledge. Hopefully
[30:53] if your analyst knows this, they know drive-thru time. What does this week mean? What do my stores mean? Right? They know the semantics, the ontology of your business. What your agent still doesn't know is expertise.
[31:07] And this is where a great analyst goes from the what question to the why question. So how do you go back and reverse construct that? This is where the context development life cycle becomes
[31:19] important. And the way we think about this is AI builds, tests and reviews context. Humans resolve ambiguity, add tacet knowledge and approve context.
[31:34] So to give you an example, we had a uh a public company customer they're in in software. Uh the AI went right through all their SQL code and found that the
[31:46] monthly recurring revenue metric was defined differently in their sales domain and their finance domain. Actually in this case both were accurate. They were running an incentive in their sales team and there was a reason it was being defined differently
[31:58] and they weren't reporting it to the street that same way. It was actually accurate. So if you're building a finance agent, a human needs to make a decision between these two metrics. Which one do I actually choose? Right? Like that's a
[32:10] human decision. This is what we call human on the loop, not human in the loop. The human is not trying to make approve every decision. The hum the AI escalates to the humans on decisions that are that only humans can make and
[32:22] we use that to improve context over time. A strong part of this is to not make this a never-ending documentation life cycle. Every AI interaction creates more
[32:36] context and harnessing this as an observability loop can create gold. And so that's the second layer of this context repo. We think of it very much like a GitHub repo for context. AI helps
[32:48] you create it. AI helps you maintain it. Humans make decisions to ensure context is accurate over time.
[33:01] I'll end with this which is enterprise context is going to need the same kinds of challenges that we needed with data right life cycle management collaboration versioning governance uh and so we've been trying to imagine if
[33:13] you're able to take the same paradigms that we have built over the last decade in data engineering and you bring it into the world of context engineering what does that look like and what would that world look like Huh? Why don't you
[33:25] show us what this looks like in action? >> Yeah, absolutely. Get it switched over. Perfect. All right. So, we have crawled in all of
[33:38] your data estate. We know how the data flows in your organization. We use context mining via context agents to document that data estate to build a great foundation. We generated an ontology. We related the physical and
[33:51] the logical. Now we need to actually build a context repository. And here in Atlin, I can use context engineering studio. If I'm a developer, maybe I'm just in cursor, maybe I'm in claude, it
[34:04] doesn't matter. You can build these context repos wherever you're comfortable via our MCP and the skills that we have to help you build them. But for sake of demo, we'll go ahead and build this here where I'll ask it to build an agent that will help me analyze
[34:17] the M context business. That's the one we've been talking about. And I specifically want to analyze subscriptions and customer health along the way. I'm trying to create this boundary around that context. So we can
[34:29] do things like bound for example the metric uh conflicts that we might run into where I calculate MR one way but if you talk to another part of our business or in another context they calculate MR
[34:42] differently. Now, for the sake of demo, this takes a little back and forth between human and AI where we chat just like a chat GPT session here. I've already gone through and done this one
[34:54] time where you'll see what ends up happening first is it scour your data state estate to help you find those assets that could be relevant for this particular agent to have access to. Then
[35:06] it starts to document those. If for some reason they're not documented by launching those context agents along the way, we might have a discussion about some metric resolution or metric
[35:18] conflicts that it found and I as a human help it resolve those conflicts as we're going through this discussion and even do things like I know this is silly but capturing your brand voice in the end.
[35:31] Does anybody know what like the catchphrase would be if you worked at Chick-fil-A? What do they say? >> My pleasure. >> My pleasure. All right, we got one Chick-fil-A go. Thank you. All the way in the back. That might be something that you want to make sure your agent
[35:43] says all the time because that's how you want to represent your company. And once you get done going back and forth building this context repository, that's where you end up with the final product that you then want to run simulations on
[35:57] top of because you need to know, is this accurate enough for me to push out to the business? The worst thing I could do is say, "Uh, I don't know, yolo." And put it out there and see if the business uses it or it answered one what question, okay? But as soon as the
[36:11] business tested it, it got out of whack real fast. We help you generate first off those business questions that could be asked through all of that context that we mined. We run those simulations so you know your accuracy level and
[36:23] where it's wrong. We'll even generate a suggestion for how you can change that context repository to uh eventually get that question right. And you continue to iterate until you're at an acceptable accuracy level where then you one-click
[36:37] deploy to something like for example data bricks and a genie genie space or maybe it's out to claw it as a managed agent or it could be out to like a deep agent in in lang chain
[36:49] but you build it now you actually need to make sure it's good going forward. How do you continue to improve upon this? Well, this was that compounding nature or the flywheel effect. We tap into the traces that your agent produces
[37:04] and we launch our context agents on top of those to mine and scour patterns to find suggestions where things aren't clear. So in this example, there was a double charge detection that uses three
[37:16] different SQL patterns. Here's what we think you should add to your context repository. If you want to apply it, go ahead and hit apply. You can't possibly have humans looking at all the traces and trying to identify
[37:28] patterns. This is where you need to use agents to improve agents and then you use human in the loop to help steer those context repositories at PK. I'll turn it back over to you to help finish this out.
[37:40] >> Awesome. So, one of my favorite things that customers say when they start with this is they say used to take me months to deploy. Now I can deploy in a couple of days. uh I would not have been I would have had to go reach out to so
[37:52] many humans and so many different teams to get this context because as a data person I don't know this domain context we're able to reverse construct that and the last they'll say when I launch this agent I launch at 70% autonomy and over
[38:07] time I'm capturing human traces to improve it to 90% 95% autonomy
[38:21] hard-coding context. You're shipping this one agent for a single use case for a s single business, a single domain with scale. This is unsustainable and
[38:33] potentially dangerous. Context must be open and portable in this world. Um, it's something that can become that unifying layer. It's your
[38:46] IP. Something I always think about is how in this world where um intelligence is getting commoditized, everyone and their competitors have access to the same
[38:59] intelligence. Uh what makes a business unique? What makes a company unique? Uh what differentiates a customer support agent of American Express and Amazon? Uh that's how you do business. That's who
[39:12] you are. Uh and so last year this time at summit uh we dropped context is king. Uh this year uh my one message context is IP. Uh keep it open, keep it yours,
[39:28] keep it portable. Um that's all we've got. Thank you so much. Um
[39:41] >> we do have one minute left officially. If anybody wants to ask questions, please feel free. Shout it out. Questions. I thought people would be shy. Come up afterwards. I know you guys are going to storm us up here. Thank you. Oh, we had
[39:53] one. All right. Got someone that's willing to talk on a mic, >> not just me. >> There we go. >> So, how do you do your evals after the context bootstrapping? I think that's
[40:06] the part where things get automated but collecting evals from the customers and then making sure that your agent responds with the right responses and measuring those responses the confidence level. >> Yeah.
[40:18] >> All of that. How do you how do you do that? >> Yeah. So the first part of that that we do is we generate a test set uh and we actually use context to generate that test set. So for example, you're shipping an analytics agent. We know all the BI that people have already used. So
[40:31] we're able to reverse construct and say here's the 10 personas that might use this. Here's all the questions they've already asked. Let me take that and simulate and create you know a thousand test question set. And then AI does the
[40:44] first job. So it basically says do we have context already for this or not. And that's the human in the loop process that we run uh which helps you actually take accuracy up from the base accuracy to the accuracy that you can ship with
[40:57] confidence on. Uh and then you would have seen this in the context repo. Uh every time you change an update context, the same simulation flow can run. Uh and so we we use simulations as a way to ensure that we can trust we can ship
[41:10] changes with trust. >> Yeah. >> Cool. All righty, that's our time. Uh feel free to come

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.