Name: Introducing the Context Studio
Uploaded: 2026-01-28T22:30:13.990Z
Duration: 41 min 8 s
Description: Introducing the Context Studio

Transcript for "Introducing the Context Studio": Hello, everybody. Welcome. Welcome. We are excited to have you at our live event introducing the Context Studio. We will get started here maybe the next couple minutes as everybody gets acquainted with each other in the chat. We'd love to know, like, where you're coming from in the world. I am located in the Dallas area in Texas. We are currently thawing out, if that's interesting. We don't handle ice very well here, so we'll we'll get through it. But it's nice and chilly for quite a while. So, yeah, let us know where you're coming from, and we're excited to get started. We'll probably get started in, like, the next one or two minutes, I would say. Let's see. Hopefully, Hunter, your Wednesday what's today? Wednesday is going well. Middle. of the up in Portland, Oregon just getting some rain, which is very normal for us. I'm a little bit bummed that we don't have snow. So Okay. Well, welcome, everybody. As people are trickling in, let us know where you're coming from in the world. Throw it in the chat. We'd love to know how far reaching we get. We also have a poll open. So if you don't mind taking a second answering our fun little question, very relevant to our content today. Oh, we've got some folks coming in. We've got Charlotte, North Carolina. Amazing. My daughter is named Charlotte, so I have a special place for that in my heart. Oh, hey. Another Texan and somebody from Sheffield in The UK. Welcome. Cool. So let us know where you're coming from. We'll get started in, like, another minute or two, and we will get going. We're talking context curation. We're talking agents. We're talking all the things. Somebody from LA, welcome. Amazing. And be sure to fill out our poll as well. We'll read those answers as we get started and, like, kinda start with our agenda. Santiago, Chile, Berlin. Wow. Awesome. So this is going far and wide. I love seeing it. Cool. Well, we are two minutes past, so we will go ahead and get this thing started. Thank you guys so much for spending time with us today. We are excited to talk all things Context Studio. It launched just today. We have an amazing launch video that is all over social media. You actually might see one of the stars from that video in the session. I was able to secure that talent. And the context studio is really about unifying this toolkit to observe, test, and deploy analytics agents. So it's really important when teams are thinking about how do we use AI in an effective but also trustworthy way. And analytics can be particularly challenging, so we're excited to talk to you guys about kind of how we think about this, but also what we've built to solve for this problem. So let's jump in. Our agenda is going to be some just quick intros of our speakers. I'll kind of run through what Hex is just in case you might be unfamiliar. And then I'll turn it over to Andrew, who is actually going to talk about this thing we think that the context studio really enables called the context development lifecycle, and then what we built to help solve for some of the challenges that I'm going to talk about, and then he's going to cover. And then we'll have Hunter walk us through a demo and kind of a live scenario and workflow in the context studio so you can see it in action. And we also will have time at the end for q and a. So as you're kind of listening and hearing and you have any questions, please make sure to put them in the Q and A tab on the right hand side of the screen, and we'll get to answer those questions live with you guys at the end. Amazing. So before we get started, let's look at the poll results. Let's see if I can figure out how to do that. I don't know. Okay. We'll skip the poll. I'll keep you guys in suspense and we can come back later. So without further ado, our speakers today are Andrew Lee. He is the product manager here at Hex responsible for helping build the context studio. Also got honey Hunter Henninger, our sales engineer, who's gonna walk us through the product. And then my name is Rachel Herrera. I'll be your host today, and I get the distinct pleasure of being an evangelist for Hex because I love it so much. Love the data space. Excited to share it with you guys. And if you're unfamiliar, Hex is an AI analytics platform. And so what we really believe in is the ability for anybody to come in and become a data person. And so we've built sort of this platform that provides AI augmented tooling for data professionals to come in and build SQL, Python, dashboards, apps, etcetera, but then also for anybody else to come in and ask their questions in natural language using our agents. We call ours threads. But we kind of have agents all over the product doing all sorts of things. And all of those agents are kind of governed and use this compounding context engine in the middle. And that's really important because we wanna make sure that agents making decisions on really deep data work or business user questions all are working from the same sort of source material. And so as you can see, HEX really thinks that AI kind of serves a lot of different purposes, but it needs sort of that grounded central context to be really useful. And when this works well, we feel like it causes or creates this thing we call the virtuous cycle, where data teams are creating context and business users are asking their questions. Maybe they get a new idea, and then they have to send that back to the data teams so that they can go and create new context, dashboards, etcetera. And when AI gets introduced into the cycle, it starts to spin really fast, which is really exciting and great, but it also creates some of this friction or at least some questions that are really top of mind for data teams today. Things like, when I'm updating this context, what should I be updating? Like, how do I know if if I pull this lever here, will it do the thing that I expect, or will it make the impact that I think? And then when business users are asking all of these different questions, how can a data professional or somebody at the helm of this agent understand what are they asking and what's most important to the business? And then I think the most important part is where are they struggling and where is maybe the agent not giving the best answer or the user getting confused with what the agent's responding with. And really, this sort of boils down to this idea of how do we maintain trust in this loop, especially when agents are sort of at every stage of it. And that's really the crux of why we felt like it's something like the context studio is really important because turning on AI is pretty easy, but making it scale isn't. And so with that, I'll turn it over to Andrew so that he can walk you guys through sort of what he built with the team and why it's so important. Thanks, Rachel. Yeah. So, you know, HEX has been working on AI features for data analysis for a long time now. We started, early on with some magic features. We have a ton of agentic capabilities today. And what we've learned from both internal testing as well as with our customers is that there's a lot of complexity here, actually, in making AIs really great at being able to work with your data. Context is one of the core problems to solve here. You know, there's a lot that's happening with the frontier agents here. They're getting better by the month, by the day. But what makes an agent really able to work with your data in particular is the context that it has. And that's really what makes a huge difference in terms of how well the agent is able to answer the questions within your organization. And a lot of different tools have different solutions for providing this context. But what we've learned is that there is no single one size fits all approach to context. There are different tools that are necessary for different aspects of tools for the agent. So there are things kind of on a spectrum of rigor. On one side of things, you have, softer guidance. So for us, we have, like, unstructured workspace guides, things like, metadata inside of your warehouse describing columns, describing tables. And all of this is guidance for the agent on how to use your data. But it's not, like, deterministic. It's not hard rails. But it's incredibly important for the agent to know about your business context, how to use your data, and all of that is embedded within this sort of unstructured guidance. There are also kind of like more guardrails that you want to propose for your agent. So for HEX, that would be things like endorsements. But, basically, you want to make clear what sorts of resources the agent has access to, like which tables, which schemas, which semantic models. And then, on top of that, you might want to have more nuanced rules. For less technical users. You might want a more constrained set of resources. For technical users, you might want them to be able to kind of go off rails, maybe work within a dev schema. And you need some set of clearly defined rules around what the boundaries are for the agents for those different types of user groups. And then on the strictest sense, in terms of governance, there's also tools like semantic models. This is a pattern that we've carried over from the BI days. But there are times when you very explicitly want very deterministic SQL generation for very important metric definitions. And you just want to make sure that every time a question is asked and answered about ARR, it's using the exact same definition but with the flexibility to still slice and dice the data. And that's what semantic models are great at. These are all very different tools, but work really powerfully when in concert with one another. And so what we've learned is that we really need an approach that combines all of these different types of tools for the agents. And that's a big part of the focus area for HEX in the past few months. And once we had those tools in place, we realized there is actually a very important development life cycle around this context. Yes, there's, like, a richness in terms of the types of context tools that HEX provides. But on top of that, there's an overall development lifecycle that also needs tooling around how you can identify and target and test and deploy these context changes. So we see this, like, general life cycle around the context, like you see before you. It starts with just, like, understanding what the agent is doing. You need some way to understand what your users are doing with the agent, what the agent is answering, and have some way to monitor for quality as well. Usually, once you get that kind of high level picture, you need to be able to take specific, test instances, like specific conversations, and really dig into and understand exactly how an agent got to that particular answer so you can figure out what the right context intervention would be. And then, finally, you get to figure out which one of these context tools is the right one to use. Usually, it's pretty obvious depending on what you've learned from the deep inspection about that particular conversation. But before you deploy those context changes, you need some way to safely manage your change. And that requires some way to test your changes before you land them. But once you do, you know, you're back off to the races, observing and inspecting and curating. And so to bring this entire development life cycle together in one place, we really built out this context studio. And so that's why we've consolidated a lot of these tools that we've kind of built over time into a single, workflow within HEX itself so that you can really develop your HEX agent in a way that is really good at answering your particular organization's data questions. So here, in one unified flow, you can observe the agent behavior. You can diagnose those issues. You can make the appropriate types of context changes and also do a test and deploy workflow. And so, really, like, taking a step back, there's two main parts of this. And if you think about the way that the cycle works, one huge half is just the ability to observe and inspect. So we have, one, observability feature where you can see volume of conversation. You can see all the specific conversations that are happening. You can see things like what the top topics were. And you can get a sense of quality as well with, different classes of different potential warnings about agent behavior. And then on the inspection half, we have a threat inspector where you can really drill into and understand exactly an agent's reasoning, what tools it used, as well as understand, exactly what triggered a warning and what a potential remedy for that, issue could be. And then once you get to the actual context development portion, we have, like I mentioned, these different tools for the different types of context jobs that you need to be done, as well as test and deployment workflows to make sure that you're safely making your changes, making the changes that you want to see in your agent behavior is changing, but also making sure that the types of agent behaviors you want to make sure are the same are also happening, all before you actually ship your context changes. And so this is really bringing together this full life cycle in, like, a very straightforward consolidated place inside of Hex. And I'm excited to hand it off to Hunter now to give us a demo of it all in action. Sweet. Let's, jump into the product here. I'll go ahead and share my screen. Cool. For anyone that is new or newer to Hex, just to orient us orient us around what Hex is before I jump immediately into the context studio. Hex is a collaborative data workspace for both data experts and data consumers to work together in the same platform. So we have a note booking and project architecture where data experts can go deep with SQL and Python and visualization creation as well as entry points for data consumers to use basic filtering, drag and drop tools, and even natural language to interact with data, which is what you're seeing here from the home page with the this ask functionality, which we call threads, which segues directly into managing context in the context studio. So I'll dive into the the studio here. What this is is the batteries included one stop shop that Andrew was just mentioning where anyone that cares about context curation and data engineering, thinking of anyone on a data team, like data engineers, analytics engineers, those types of people can manage how well context is being, curated and, received by agents throughout HEX. Hex has a lot of entry points for agentic analytics, and this is your place to get a high level overview of what's going on and and make improvements, proactively. So the context studio, the way to think about it is broken up into three main workflows, the first one being Observe, the second one being Inspect, and the third one being Curation, all things that Andrew was just talking through in those slides. And I'll kinda go through those one by one and how a user would interact at each step. So first, with observability, when you land in the context studio, you can see this nice dashboard view of all agentic conversations happening in hex from a user volume perspective, an agent volume standpoint, and also emerging themes in what we are calling topics. Below these visualizations is a full list of all agent conversations happening inside of Hex with a bunch of metadata associated with each. So we're we're telling you which agent it was, which user it was, the role, plus some extracted information, in the form of topics and warnings. As I scroll through this, you can see some examples of these where we are calling out, hey. There is a theme of questions coming in related to a certain topic that we are deriving from all of the questions that are being asked as well as any room for improvement in the form of missing context, user doubts, potentially some agent confusion in the form of, like, a caveat. We're gonna flag those things to you so that you can proactively make improvements at the context level. This is a very filterable view here. So you can see one simple thing that I can do is, like, zoom this out to the last thirty days instead of the last seven days, and we'll see these themes kind of change live. I can also filter this by agent role. Maybe I wanna dig into a specific user's interactions. But a really common scenario that that we are envisioning this playing out is at the topic level. So if I'm looking at this chart here and I see that enterprise segment analysis is an emerging theme in the last thirty days, but it does have a relatively large amount of warnings associated with threads that, people are asking about. Something that I can do is come up into this and filter to just the enterprise segment analysis threads and get a list of all of the questions coming in so I can get, like, an initial glance at what are the questions happening, and and how can I learn about ways to improve this? We can see that this top one where I was preceding this demo with a question about what our net dollar retention is for enter enterprise customers has a few warnings associated with it, a couple times of missing context that the agent is calling out, and the user, itself actually doubted the response that the agent was giving back to it. This is a great and and valuable moment as a workspace owner to say, like, great. I understand that all the answers related to enterprise segment analysis are not perfect, and I can go and prove this in a very actionable way, which moves me into the second phase of inspection. This is where I can click into a specific thread that I want to investigate and get a an inspection workflow of what happened, where was the user doubt, where were the warnings kind of tied to this thread. So I'm seeing the thread here on the left side. These are all of the agent interactions and user interactions as the someone worked through these these questions. On the right side, though, I'm getting a quick glance at a summary of of what question was asked, what was the response, anything to call out that the agent wants to mention, plus a timeline view that is a little bit more specific. I can understand at the question level, because sometimes threads can get pretty lengthy with many follow-up questions, What was the question and what was the agent's approach? As well as all of the steps that the agent took, all of the tools that it called to understand the context initially, what queries did it write, which semantic models did it access, all of that good stuff. These are all clickable. It'll scroll me to that part of the thread where this tool call was made. In this case, the second question actually expresses some user doubt, and this timeline tab is actually calling this out to me, which is great. I can scroll to the question where the user was expressing some doubt, where I, as a user, said these numbers seem a bit off. Are you sure about this? Great context for me to move into the third tab in this inspection workflow, which is warnings and an expression of which what the warnings actually meant in this case. So there was a lot of confusion from the agent on how to calculate net dollar retention. It felt like the number was a little bit low in this case, and then the user expressed some doubt. The best part about what Hex is doing here is we're actually providing you with a very actionable suggestion to go make a change at the context level, whether that's adjusting some metadata in the warehouse, adding more columns to the data connection, adjusting some metrics at the semantic model, or just adjusting some context in the form of those guides and rules files that Andrew was mentioning. So the the way that I would move from this step is go back to the context studio and make some tweaks, which I will do here in a second as kind of the third phase, which is curation. But I do wanna pause here and see if Andrew has anything to add in terms of Hex's approach to building this content and and anything that you'd like to include in terms of why we went the direction that we did. Or is Andrew not there anymore? Andrew, Nope. you there? Sorry. Lot lot couldn't find the unmute button for a second. Cool. Yeah. I think, to be, like, frank, like, a lot of what we've built here is in response to, like, our own and our customers' experiences, trying to understand how agents are getting to answer. So a lot of this has come out of our own, like, kind of deep engagement with these agents. Prior to this, we were doing a lot, both using an internal dev tool, whose nickname is Shoebox, that has a bunch of, like, deep dev introspection tools as well as simply, like, pairing with our internal data team and just reading all of these threads together. We would literally, like, read all the generated SQL queries together. We would read all the thinking messages together, sort of stack hands and, like, synthesize what we understood, and then figure out the next steps. The dev tool that we were using internally also just had a few additional things to introspect, like the tool calls and exactly what the sequence was and what the timeline of what the agent was doing. And we started to, like, sort of lift these patterns up into the product itself, for, like, the kind of, like, shared synced readings where we would read the threads together. That's where things like the overview and the warnings came out of. And from, like, the internal dev tool, that's where, this, like, kind of, like, really detailed timeline came out of as well. Cool. Thanks. Super helpful context. We'll move into the third phase of the demo, which is curation. I can take these suggestions and and take action on them and even test the results of me making a change before I'm actually publishing and and making things go live for for all of my users that are continuing to ask questions. So I'm gonna go back to the context studio. And over here on this left side, you can see under the management section, there are a few entry points for context curation in Hex from semantic projects that we can inherit from things like Snowflake semantic views, cube, metric flow by DBT. You can also author semantic projects directly in Hex through our own YAML spec. Endorsements is where you go through your data catalog and make sure that the correct schemas, databases, and tables have your stamp of approval from a Workspace owner perspective so that the agent knows which schemas to, prioritize, which schemas to ignore. This is your your way of nudging the agent in the right direction towards the the golden layer of tables that exist in your data catalog. And then we have guides, which are more of a raw markdown style context. We have one called the workspace rules file, which is something that the agent will always refer to when answering questions. What's in this rules file is always true information because it's always going to be referenced versus in the guide library below. Similar concept, but you can establish these on, like, a use case or a concept basis that are dynamically retrieved by the agent when it's identifying that a certain topic is being asked about. So in this case, the user was asking about net dollar retention, and there was a bunch of confusion both by the user and by the agent. So I want to come in here and publish a guide that steers everyone in the right direction and make sure that these calculations are mapped to how we think about things internally. When we go to edit this, you can create as many of these guide files as you'd like. The one that is currently published is one called churn, where we're telling the agent, like, which tables are meaningful from a churn perspective and how we define this internally. Something similar is here for net dollar retention, but this is unpublished. This doesn't exist in the workspace yet until I go through and, firmly publish it. I preceded this before this demo so that we can kinda go through the flow here. But you can see that I'm I'm clarifying a lot of the things that the agent was confused about here in this guide. And the best part about this is that I can go through the process of testing some threads on this before I'm actually making the change just to confirm that, one, it looks at the guide, and, two, it's spitting out better answers than what I was seeing before with some of the confusion with my users. So you can see that I'm adding this brand new guide file. It's showing me the the diff view here. And then I can go directly into a testing threads workflow where I'll put in the question that the user was asking initially. And what I'm looking for here when I hit enter is is the agent referring to the guide at the first step through its thinking process so that I have the peace of mind that it is reading through this context that I'm spitting out? So you can see that it's it's looking at all three of these items. It's looking at raw tables in Snowflake, modeled data within our semantic layer in hex, and these guides. So it's actually identifying that both net dollar retention and churn could be related to this question. It's gathering as much context as it can, and then it'll go through the process of answering this question from scratch. So you can see that linear flow where you go from observability, of identifying emerging themes, into inspection of places where you feel like there is room for improvement, adjusting the context from a curation standpoint, and testing the results. And then it turns into a flywheel effect where once I publish this change I'll go ahead and publish this because now I know that it's referring to that. I can now go back to the first step of observability and see if the questions related to enterprise segment analysis start to improve from here. This is something that we'd expect teams to look at on on some sort of, like, weekly basis. We've been hearing from some customers that people are so excited about this that they're looking at it every day. I think it's a really exciting entry point as people are starting to use more conversational analytics on top of data, especially in HEX. So that's all I wanted to cover. We're we're happy to answer questions, but I'll pass it back to Rachel before we do that. Awesome. Thanks, Hunter, and thanks, Andrew. So we do have a nice stream of questions coming in, so let's jump into them. I think I can highlight them, but maybe not. So I'll just read them out, and then you'll have to remember them. Let's see. There was one that had a lot of likes on it, and I think it was one of the first ones to come in. Yeah. So question is, is there a path to have this work sort of like the skills files owned in a Git repo and pushed to HEX workspace roles? Things like Cloud, Copilot, etcetera, are used today. So I I think, Andrew, this is something that comes up kinda often about, like, the interoperability of some of these guides. Like, do you wanna speak to that? Yep. Thanks for the question. Yeah. This is something we've gotten as feedback, quite consistently on this. Like, is there any way that I can just push, a bunch of, like, files from my Git repo, to programmatically update guides? And the answer is not today, but it is something we will be working on pretty much as an immediate follow-up. Yeah, like, everyone wants this. We want this internally. So this is definitely coming soon. I've got another question. Is there a way to create custom warnings? Like, if I don't want end users to use AI for, you know, specific kinds of questions. You know? I think custom warning is also something we hear or we've heard recently. Yeah. This is a bit, I would say, like, lowercase w warnings, like, not the warnings feature we just just demoed. But in general, if you want the agent to respond to specific topics, with, like, some sort of deflection to the data channel, you can put that I would I think the safest place to put that is actually in the workspace rules file. You could also put it in a specific guide. But if you just want, like, extra guarantee that any time the agent hears about this, it will deflect to whatever behavior you tell it to do, you can definitely put that instruction in a guide. We actually had a case of this internally where, for a specific, topic that the data team just knew was being asked about quite often but didn't have model data ready yet, there was an instruction that whenever someone asked about this to redirect the user to talk to the data team in their Slack channel, and that worked great for us internally. Let's see. Another question. This there a way to identify misplaced confidence where something about the context is making the agent think things are clear even though they aren't? Interesting question. is an interesting question. At least from the AI research that we've done so far, LLMs evaluating themselves in this sort of manner just doesn't work very well right now. The LLMs can't, like, judge each other in this specific sort of way. Like, if the LLM was confidence once, it'll not, like, sort of just think harder and find itself to have been overconfident in the past. So this also yeah. I agree. This would be great. We're still trying to figure out a good pattern for this, but we haven't found this to be, like, practically ready in terms of, like, the way that the models work today. So we still have to rely on humans is what I hear. Yeah. Yeah. That's, like, explicitly, like, why we call, like, the warning, feature, like, warnings. Like, the best that we can provide with, like, LLM as judged today is reading, like, the thinking messages and the user messages and finding instances where it's kind of explicitly called out either, like, there was missing data or, you know, there's oftentimes very clear instances you'll find where the user pushes back on the agent. Like, those things are really easy for, you know, the LLM basically as a form of summarization to extract those terms in the conversation. But for an LLM to kind of look at an LLM generated analysis and just, like, find the gaps, it just is not quite there yet. The other pattern that we're thinking about that can be helpful for this sort of thing is an eval suite. That's something we're thinking about in the near term as a next project where you can sort of predefined prompts that you can run on a schedule and also define how the agent should have answered the question and just be able to predefine these tests, run them on a schedule, so that you can make sure that you're not at least regressing on important tasks where there's a clearly correct answer or approach to analyzing the data. And Sid had a question. They asked, with the new guide feature, I see a world where the agent can audit existing and analyst creative reports for accuracy and flag reports with potential issues. So I guess maybe the spirit of this question is, like, kind of retroactively applying some of this against some of the stuff that's been created and then maybe, like, being more proactive in the act of making it. Like, what are some of the thinking, or what's some of the thinking there? Yeah. This is a really cool idea. I'll say taking a step back, like, a lot of this, you know, kind of the demo and the way we're talking about a lot of these, like, context curation features is really focused on, like, a self-service pattern. But the way that all the agent, whether it's operating inside of a notebook or inside of, like, our more, like, self-service interfaces, they all have the basics, like access patterns for all of this unstructured context the same. So they all get the workspace file inserted on every conversation. They all get the user preferences inserted on every conversation and then dynamically retrieve additional guides using a retrieval tool based on, like, file name and description. And we're kinda we're still tweaking the search algorithm, but, the guides are dynamically retrieved. That's all to say that you could sort of test this today by simply using our notebook agent. So you can go to an existing project that you've developed, use the notebook agent, and have it leverage the guides and have it inspect the Python and SQL that's already been written and see if there's any conflicts with the guidance inside of the guides. Very cool. We've we've got a lot of questions, so thank you, audience, for being so engaged. I think one that I saw that was interesting was it's probably this is a bit of a softball, Andrew, so I'll lop this one at you. So Justin asked, is it possible to download agent user conversations to analyze outside of HEX? This is not possible today, unfortunately. We have heard this request a few times, though. So it's something that's been in our backlog we haven't quite gotten around to yet. But, yeah, I think a lot of people have been asking for this. Gotcha. Okay. I think we'll have one more question, and then we can wrap it up. William asked just a minute ago, with one of the warnings being miscontext, have you guys explored the option to offer the approach that many Angensic tools are taking with the planning execution where the planning phase can interview the user for more context upfront? That's interesting. Yep. Yes. This is something that we've talked about internally. Like, obviously, we're using a lot of the same coding agents that you all have. Like, Claude Code, I would say, is particularly good about interacting with you as a user to come up with the right execution plan before it goes off and, does a bunch of stuff and edits half your repo. So, yeah, we're thinking a lot about, like, what the right set of tools to expose to the agent are. I can imagine a lot of this planning and execution, stuff would be very helpful. We're also experimenting with, like, faster and slower modes where the agent has more time to plan before it executes. So that's all to say. Yes. These are things we're thinking very deeply about and, is in, like, mostly r and d phase that we're testing internally right now. And I'm I'm getting word from control that we can have one more question. So if you're up for it, are you thinking about supporting custom Rag connections, vector stores as a tool available for the agent to use based on context, things like OpenSearch, MCP, etcetera. Yeah. I think the one that we're thinking about right now in the short term is MCP. It's obviously, like, the most well known and well integrated with a bunch of different sources. So right now, Hex does not operate as an MCP client, although we do have an MCP server. So we're thinking about also including the ability for Hex to be an MCP client, which would open up the door to external sources of context that are dynamically retrieved. This kind of, we think, like, kinda complements the other approach that I was talking about earlier where you can have, like, git synced context that's more static and kind of, you know, kind of like the bedrock context that you won't that you know will not be shifting under your feet. In addition, you'll have these more dynamically retrieved sources of context, whether it's, like, a data catalog, or we even have customers who have built internal agents for specific use cases that they want to expose to the hex agent via MCP and what you know, it's whatever tool is exposed via MCP that can be dynamically retrieved. I think once we are planning on supporting both patterns and what will make everyone feel very safe moving forward to, like, implement and update these is what that eval pattern that I mentioned as well. Wonderful. Well, I think we got through most of them. If there's a question we missed, we will follow-up with you after this event. Thank you, Hunter and Andrew, so much. I guess I'll ask the kind of the closing question. Hunter, what's the thing you're most excited about with the Context Studio launch? Like, how do you feel like it's gonna get received by the individuals you speak to most frequently? Yeah. I think it's gonna be a really exciting thing to work with our customers on. I think it gives us a really nice entry point to evaluate how threads are being received by users without having to, like, go ask everyone individually. It gives us a really good, like, eagle eye view at what's going on, especially as people are evaluating Hex and thinking about expanding it to more users internally. I think it's gonna be a really good tool for that. So that's what I'm excited about as a sales engineer, as you can imagine. Of course. And, Andrew, anything particularly top of mind for you? And you've you've lived and breathed this for quite a while. Yeah. I mean, the thing that I'm excited to tackle next is now that we have this observability surface and we have all of these context tools, reducing the friction that you can additionally, like, kinda compound that context like we were talking about. But, practically, for me, like, if I was on a data team, I would want the agent to be, you know, kinda pushing suggestions to me that I could quickly review, have an eval suite prerun on where I could just almost, like, be like, Okay, yeah, this looks like a good PR to me. And I accept it, and I just move on to the next suggestion. This sort of, like, very low friction with human in the loop workflow is what I'm really excited to start focusing on as an extension of what we've already built. There's, like, a hint of that with, like, the warnings and the suggestions, but there's kind of like aggregate higher level tools that can help you coordinate the changes across the context tools that I'm really excited to start tackling. We're already talking to our data team internally and, like, kind of going through and proposing workflows. Because we know oftentimes, one friction right now is coordinating changes between the guides and the semantic models and maybe even upstream into your DBT metadata. And we're excited to kind of, like, push into, like, a more cohesive workflow that really works well for data teams. Amazing. That sounds so exciting. We gotta start somewhere, though, it sounds like. And I think it's a pretty darn good start. Pretty excited. Amazing. Well, thank you all so much for spending the afternoon, evening, morning, wherever you are in the world with us. We will share this recording out. So if you'd like to share it with friends, family, whomever, please feel free to and stay in touch with HEX on all the socials. But until next time, you all have a good one. Thanks, everyone.