Agent Bricks: Building Multi-Agent Systems for Structured and Unstructured Information

Key Takeaways

Hello Brricsters. Hello. Hello. Welcome. Welcome. Welcome to Summit 2025. We're so excited to have you all here. All of you streaming at home. All 65,000s of you and so many of you in the room here. So many of you here in San Francisco with us. We are just super excited. You're going to hear all of us say how super excited we are. We're going to build off of the momentum of all of those exciting announcements that we heard this morning at our keynote. And before we do that, we want to take a moment to introduce some of the things that you can look at to uh if you're here today on campus on Wednesday and SF with us. So, please feel free come downstairs, visit us in the expo. We have so many wonderful things going on downstairs. We've got sponsor booths that you can come visit. We've got all of our partners, our amazing, incredible partners here with us. We've got brain dates that we'd love for you to join us in those. That's something new that we added this year. And coming to you live from the room here in 208. If you are watching at home, we've chosen some really incredible speakers for you to listen to. And perhaps, you know, maybe you're in the hallway and we'd love for you to come in. We've got plenty of space here. So, we're going to kick it off by hearing from our first speaker and Ely is going to again build off of all of the exciting announcements that we've heard this morning. We're going to talk about agent bricks. She's going to tell us how we can incorporate human in the loop feedback, how easy it is, how we can democratize data for everybody, work in our our vector search, and make it super easy for everybody at home to build these agents. So, with that, we're going to turn it over to Elise. take it away. Awesome. Perfect. Thank you for the warm welcome, Allison, and welcome to all of you. Thank you so much for taking the time. Uh, as she said, we are super excited to have you here. If you signed up for a talk called Talking to All Your Data, surprise. We're actually here to talk about agent bricks, but I think that's a really relevant topic. How many of you have actually wanted to talk to your data? Yeah, that makes sense. I Me too. Uh, obviously I love data. Uh, even in my personal life, I think this can be super applicable. So, I'll give you a quick anecdote here. My dog, her name is Fern, and I do a dog sport called barn hunt. Uh, the TLDDR of barn hunt is that you put rats in PVC tubes. They're well protected, totally fine. Hide them in some straw, and then your dog tries to find them as fast as possible. Uh, funnily enough, my dog's actually really good at this. She's nationally ranked. we've gone to nationals and placed fourth. And so at this point, she actually has hundreds of results. And so I really like to use Genie to be able to see how she's doing over time. But what I'd love to be able to do is be able to take the rule book that we have, understand how titles work, and then combine them together so I can talk to my dog's data. And so that's what I'm going to show you how to do today. I won't show you specifically a barn hunt example. it'll be a lot more relevant to uh your business and enterprise. But that's just the start of the possibilities we have here. If you've been to other sessions before, you might have seen this slide. Essentially, a lot of this is forward-looking. Uh maybe some things that we mentioned that are not released yet. So, please take that do not take it as, you know, uh the word of God at this point. So, uh it's definitely something we're intending to do, but things may change, especially as we're looking to shift And please complete your surveys. We take feedback really really seriously. It has a direct impact on the content and how we organize data and AI summit. So please take the time to complete your surveys. We really really appreciate it. So what are we going to cover today? We'll start with the state of the union. Why did we go endeavor on building agent bricks? What was our motivation there? And we'll go into an overview of what agent bricks is. I know you just saw some of that in the keynote, but we will dig a little bit deeper into some of that. We'll then use knowledge assistant for unstructured data, start to build out a multi- aent together. We'll use Genie for structured data, and then finally, we'll bring it all together with our multi- aent supervisor. All right, let's start with our state of the union. Why did we go and build agent bricks? This is some examples of some of the things that we've been hearing. I don't want to put something in front of my customers unless I know it's working and I know it's working over time. How do I effectively measure what I put out there? How do I make sure it's meeting the metrics and the quality bar that I expect and that it really gives the best impression of my company to my customers or even to my internal employees? There are so many different pieces I could use here with new announcements daily. What do I change? This is something Ali mentioned as well during the keynote where there are so many knobs. There are so many new research techniques coming out. There are so many ways to change and improve your agents. How do you know which one to use and how do you know which one's going to be effective? And finally, newer models are great, but do I really need it for my use case? We all know that the newest models are really smart. They can be really fast, but they can also be quite expensive, especially depending on what you want to do. So, is that the right choice for what you're doing? And how do you make that choice? This really all leads to asking yourself, is this worth the effort? So, the first question here is, how do you adapt AI systems to your needs? As we've talked about, LLMs do not understand enterprise data well enough. They have a lot of general intelligence, but they don't have intelligence about who you are as a company and how you operate, some of the terminology that you use. There are also a number of different tools and components to research and integrate. You can see that with my example on the right here. This is just a a small subset of all of the different tools and components that you can use to build agent systems today. And it's growing every single day. How do you get to the right quality? A lot of folks do not have labeled data. It's pretty hard to get highquality labeled data and it's especially hard to do that at scale across many different use cases. So, how do you then measure and make sure that you're actually meeting the mark? And then finally, does the ROI make sense? Larger models again are really expensive. Will the end users keep using the product? Does the output quality justify the cost? These are just some of the questions we've been asked by our customers and why we've spent so long uh building agent bricks. And so I'm really excited to introduce that to you today. Introducing agent bricks. With agent bricks, you can build agents autooptimized on your data. The first thing that you do is you specify your problem. So we have a number four different use cases that we'll look through today. So you give us your data and you tell us the task in a declarative way what you would want to go accomplish. We then optimize on your enterprise data. So we build you the best agent system on a co quality versus cost trade-off. So, how do we make sure we're helping you achieve your goals depending on what the use case is and depending on what you and your business need? And then finally, we incorporate continuous improvement. This is something that you heard about from both Ali and Hanland and Casey, but we've been working really hard on something that we're calling agent learning on human feedback. Ah, LF. I'm just going to call it agent learning for my own simplicity. But this is how we start to incorporate that natural language feedback so that we can continue to improve the agent system that we're building for you, not just shoving things into prompts. This is all backed by the Mosaic AI platform that we built over the last year and the framework that we've built. And it's also backed by our in-house research team. And if you're around, some of them are actually up here at the front, so you can always ask them questions. but they've been working really hard to research to incorporate new techniques to keep up with the technology and all the news and really make sure we're giving you the best of what's coming out in the world and what will make sure that you actually achieve your goals here. So, how does agent learning work? So, I know we walked through this example probably about a half an hour ago, but I think it's really important to solidify what this is. So in traditional ML you would have uh your model give an output which then produced a numerical score and so it would then learn from that numerical score especially as you changed weights and um so it was really numerically based and so this is what the model understood. This was its sort of natural language. So with agent learning what we've been able to do is learn from natural language guidance. So how can I have a model that produces an output where I then get a score? So potentially a thumbs up or thumbs down or even a grading system on an A tof basis based on your experts. But then how do I also incorporate natural language? And that's really something that the LLM understands. It speaks natural language and so it needs feedback in natural language as well to be optimized in the right ways. This really allows agents to be more precisely steered and improved. So this is how you get control over your agents. you understand what they're doing and you can really build powerful systems end to end. Let's take an example here. So in my agent system, I have a number of different components. I have my prompt. I think everyone's familiar with that. I may have a vector search index so that I can look up information. I have my LLM judges so I can see how I'm doing. My evaluation data set. I may have MCP servers with tools and even my agent configurations. So, there's a lot going on here and really up until this point, our main lever has been updating the prompt. And I don't know if you're like me where you've ended up with a prompt that's really, really long and nonoptimal. And then the more that you add into it, the more chaotic it ends up getting. And so, there are more components that we could improve here. And that's what we've spent uh a lot of time researching and figuring out how to do for you. Let's say I want to adjust the answers to ignore data before May 1990. In this case, agent learning automatically adapts the agent to that feedback. So it may adjust the prompt guidelines. We may filter and update the data so that it's a lot more recent. We would add in an LLM judge so that we can understand the recency of documents that are coming back and make sure that we're really meeting that guideline as well. could add guidelines to tool descriptions in your MCP server or even optimize with new configurations. So there are lots of ways to improve the system and we've really worked on incorporating that with agent learning. So now I'm going to introduce you to our use case. This is not going to be my dog's multi- aent system. We're going to build out a telco customer support multi- aent together. So I'm going to have a supervisor. It'll route between different agents below and then I'll have a number of different agents across both structured and unstructured data. So the first three here you see my account agent that pulls my customer information. It has their churn risk. It also has uh what plan they're signed up for. I then have my billing agent which has billing and usage information. So what has the customer been doing over time? Do are they late on their bill? Was the last one failed because of payments? Information like that. Then I have my product agent. What this does is it pulls plans, devices, and promotions. So it knows all about the devices that we have available, the promotions that are available, as well as all of our plans and different things like data caps. And then finally, we have our tech support agent. This is really a Q&A for technical support. So, it's taken in all of my knowledgebased articles and support tickets so that I know how we've solved issues in the past and common FAQs and even some of our uh common policies that people ask about. So, the first thing that we're going to build here is with knowledge assistant. We're going to build that technical support agent that can answer questions over my knowledge base and about policies at my company. Knowledge assistant is designed to democratize access to information with higher quality rag. It gives you easy access to unstructured data. You can leverage docs, powerpoints, text, markdown to give users all the answers they need. It also has built-in metrics and evaluation. That's really important as you're starting to build these systems out. How do you make sure that they're actually achieving the quality bar that you're intending? And finally, it's production grade. I know that there are other one-click rag solutions on the market. That's really not what we're targeting here. We are really intending for you to spend your time on improving quality. And so, we don't want you to spend your time building out the different components. That's what we'll do for you. And then we'll work together as we optimize on your customer data and we can achieve really high quality out of the box. So, we compared to a couple of other cloud agent uh essentially rag products on the market. I'm not going to name names here, but we are finding that Agent Bricks is significantly higher quality out of the box. We spent a lot of time figuring out how to parse files, how to do chunking and embedding in the most effective way so that we can give you the best answers directly out of the box. And so you can use knowledge assistant for the answers you need. End toend ingestion is handled for you. You can also use a pre-existing vector search index. So if you've already invested in that, you can bring that into knowledge assistant and still take advantage of agent learning. And again, agent learning and evaluation are built in. That's really important here if you do want to improve quality and continue to see that quality rise over time. And then you get a ready to scale endpoint out of the box. So you can easily embed your endpoint wherever you need it, including in a data bricks app. We actually have a a new template available on the marketplace. So if you do build a knowledge assistant endpoint and you want a really nice highquality playground-like experience in an app, it's directly available for you. The quality also improves over time. So I'm going to explain a little bit of what's going on in this graph. So what we're measuring here is our SMMES are giving feedback on an A to F scale. So was this a good answer or a bad answer and then giving guidance as well. We actually run a set of evaluations against our product on a very regular basis across a couple of different corpuses of documents. One of them being finance bench. A lot of you may be familiar with finance bench. A lot of people sort of rank themselves against that. But we also use something we call docs Q&A where that's actually the data bricks documentation. So how good of answers can we get out of that? And we like to run this because we really want to make sure we are building a high quality system that we're testing across different evaluation sets. And so what you're seeing here is that even with just two or four or eight rounds of feedback, we see significant improvements in the quality over time. So this isn't a case where you need to go label thousands of examples in order to do something like fine-tuning. We're able to actually improve the system with a low amount of high quality information from your thememes so that we can improve this. And I think that's really powerful. It's really hard to do fine-tuning and it's also hard to know if something's objectively correct. And so it starts to incorporate what you know about your business and what this means for you. So take an example like lifetime value. For different brands and different companies, lifetime value means something really different and so does churn risk. If you're a retail brand, you may want someone to come back and purchase within three months. But if you're a car manufacturer, they're probably not going to buy another car in three months. So what does that churn look like? and what does that mean for your business and how do you actually incorporate that into the model so it knows more about you? This is a quote from one of our private preview customers, Patrick at an Analytics 8. He may actually be here. Um they leveraged uh agent bricks and achieved a 40% increase in answer accuracy and an 800% faster implementation time across a number of different use cases. That may seem like an exaggeration, but they were putting in a lot of work depending on the use case to use different vector search indexes, different chunking and embedding strategies. And so not having to worry about those components has saved them a lot of time where they still achieve really, really high quality results. And as you see at the bottom too, post launch, we've also observed that answer quality continues to climb. So that's really important here. So now let's jump in and actually uh look at what knowledge assistant looks like. So I'm going to jump into the data bricks platform here. I'm going to go over into my lefth hand panel and down under what used to be machine learning. We have now renamed it a IML. There's a new tab called agents and I'm going to click there so I can see my agent bricks. As I mentioned there are a few different use cases that we support today. So information extraction if I want to take key information from unstructured data and turn that into JSON that's available here. This have custom LLM. So if I want to do something like sentiment analysis or deeper summaries I can also do that. But what we're going to focus on right now for our demo is knowledge assistant. So I'm going to jump in. I've already built this out so that you don't have to watch me do that. And so it's really easy to get started with knowledge assistant. I just need to give it a name. I also need to give it a description. So what does this do? How do people know what to ask? What types of questions to ask? Then I provide my knowledge sources. So in this example, I have a knowledge base where it may have support articles as well as FAQs and debugging tips. But I also have my support tickets. That way I can see resolutions to issues and troubleshooting steps and be able to surface that to my customer support agents much faster than they were able to search before. And on the right hand side we can see I have my deployed agent. So I have my model serving endpoint which is ready to scale with you. I have my ML flow experiment. So I can get my real-time traces as well as look at my judges and see what evaluation criteria there is and I can see the status as well of my files. So let's jump in and ask a question of knowledge assistant here. So let's say uh can't make calls but data oops works. So let's see what knowledge assistant comes back with. What's really nice is we do give you this thinking box. So you can start to see the reasoning behind the model and see why it starts to make some of those decisions. Does move pretty fast, but you can see that it realizes that both the knowledge base and the support tickets could have relevant information here. And so it's going to search across both of those in order to give me a more comprehensive answer. And so it may be due to Vol LTE. It found that within a few different support tickets. And so it can give me some debugging steps right away. Now let's also imagine maybe we put this in front of a customer. They decide to ask whatever sorts of questions they want. Um is iPhone or Android better? Let's see what our bot says in that case. All right. Right. What does it think here? It's subjective and depends on various factors such as user preferences and specific needs. So it starts to support to search our support tickets, see what people have said about the both of the devices here. And so it comes back with an answer that largely depends on personal preferences and priorities. But especially as a telecom company here, I really want to be neutral on the topic of device manufacturer. And so I'm going to go in and add a couple of guidelines for my uh knowledge assistant here. And then we can see how it directly improves right away. So I'm going to jump to improve quality, which is also right up here. And so I could generate questions, but I have some in mind. And so what I'm going to do here is go into is iPhone or Android better? And then I have a couple of different options in here be neutral. I also want it to show some of the be sure to highlight the positives of each. So I really want obviously I want them to sign up for my telecom company. I actually don't mind which phone they use. And so let's add that question in here. So I can see in here I have my question. I have my guidelines. And if I were to start a labeling session, I'd be able to give that to my experts and uh actually incorporate their feedback in here as well. Even something like good is helpful for the model to know it's doing the right thing. So let's jump back into our playground experience and let's regenerate this answer and see if it started to take into account some of the guidelines that I just gave it. Depends on does say now it wants to give a balanced answer. They're known for seamless integration. So iPhone positives and Android positives. So it did take into account my feedback right away. And you can see how that answer morphed from when we first asked it. Great. Now I have a tech support agent built out. Let's start to build out some of the other agents in here as well. So I'm going to jump back into my slides. Next, we'd like to take advantage of some of the structured information I already have in data bricks about my customers so that I can dig deeper into their issues and start to correlate across these different genie spaces. Genie is really great at helping you talk with your data. It gives you conversational analytics so you can ask questions in natural language and get answers immediately. It's really AI tailored to your business. So it incorporates unique context for your data as well as from the platform at large. So some of the things we have coming are learning from the platform metadata. So what are the popular queries and what are the popular tables? It's governed and secure. Fully governed data with results secured by your Unity catalog policies. That's important for all of these is that we make sure we're maintaining that governance over your data and someone doesn't inadvertently get access to something they shouldn't. And finally, it's accessible anywhere. So embedded in custom apps via APIs or pair it with AIBI dashboards. You can really use Genie the way that you need to for your business. And so, as I mentioned with Genie, you can leverage context from platform metadata. With Genie, you do want to pick concise topics. So, that's why I'm going to have a couple of different Genie spaces here. That way, I can keep it narrow and keep Genie really smart about that type of data. We can also enrich understanding with space level semantics. So, authors can fine-tune each genie space with example SQL text instructions. Um, and so I'll show you some of that in a second here. And finally, you can iterate based on user feedback. So your end user feedback is tracked in monitoring and you can evaluate your spaces with benchmarks. So very similar to agent bricks, you have this built-in evaluation and monitoring where you can make sure Genie is giving the right answers over time. What are customers saying about Genie? Genie now enables colleagues across trading, product, and acquisition to access data in seconds instead of hours or days, which then frees their specialists for deeper insights and predictive outputs. So, it really starts to democratize information in your organization and make more and more people more data driven and more effective. Let's jump in and see what Genie looks like. I'm going to go back to the data bicks platform. And when I navigate here, you can see under the SQL section, there is a genie. So, I'm going to go there. And I already have a couple of genie spaces built out here. Let's jump into plans and products. Let's start our warehouse. That's important for our demo here. Uh, so this one contains all of the product, device, and promotion information for a for my telecom company. And so what's really nice is you see some example questions in here. And so those are actually pieces that I configured within Genie. So I wanted those to show up for my users when they first come to the space and know what kind types of questions to ask. But I can also do something as simple as explain the data set. So what tables are there? And can you tell me how they join together? So I can see in here I have three different tables, devices, plans, and promotions. And then it also tells me how those tables join together. Let's look at how I set up this Genie room. So within context, of course, it starts with data. So you pull in data directly from your Unity catalog. If you do have metadata in there, including descriptions, field descriptions, things like that, we do leverage that immediately. So it's not something you need to add in. We also have instructions. So these are some general instructions to help Genie understand your organization and business better. So as an example in here uh I have one called promotions are available to existing customers if the description does not contain requires new line activation. That's not something Genie would know out of the box because that's something specific to how my data is structured. And then finally I also have example SQL queries. So here I have what plans require contracts. So this is another one that I wanted to be able to give it the right SQL answer so that if someone did ask this question it had a direct example out of the box. Then within settings I can provide a description. This will be important as we get into the multi- aent supervisor because it needs to know what this Genie room is about. And I also have those sample questions that help someone ask questions of Genie when they first access it. We have monitoring here. So I can see all of the questions that people have asked if they've give a rating or a request as well as their comment. Here you can see a lot of this is being used by me and my multi- aent. So there's not a lot of feedback in here, but as people use Genie directly, you can see that surfaced. And then finally, you do have benchmarks as well. And so I have a couple of questions in here. One of them being what promotions are available for existing customers. So that takes into account those instructions that I had as well as which plans require contracts. And so both of those are things that I want to test against over time. Awesome. So now we have our structured data. We also have our unstructured data. So how do we bring those together? That's where our multi- aent supervisor comes to life. This lets you design an AI system that leverages Genie for structured data. It leverages agents. So this could be for example in this case it'll be knowledge assistant but also if you built a custom code agent you can leverage that as well and you can also leverage tools. So if you've used UC functions to build out tools you can leverage those here. You'll also be able to leverage MCP servers. So I know Hanland covered that that we do have some exciting announcements around MCP and that directly integrates with agent bricks here as well. It really makes your possibilities endless. It's flexible for your use case. So, I can reason over all of my data, structured or unstructured, and then I can directly take action. That's where tools and MCP comes into play. It also has agent learning and evaluation built in. So, when you gather feedback from your experts, we can then use that to improve the model over time. That includes the answers that the supervisor is giving and the type of routing that it's doing. So how do you sort of improve that system over time? And finally, it has governance and access built in. You can leverage existing permissions with each agent. So you only expose the data they have access to. We don't inadvertently give access to something just because you put it into a multi- aent system. So those permissions really do carry through all the way. What are some best practices when you're designing a multi- aent system? One, you want to build topic specific genie spaces. I don't recommend adding all of your tables into one genie space. That's a lot for it to understand. I'm sure that's a lot for you as an organization to understand as well. And so genie is much better if you give it a specific topic. So in my case, I have one built around billing. So how is billing done? What do people owe? How much data have they used? You should also iterate on the quality in each sub aent. So both agent bricks and genie have quality loops built in. The supervisor quality increases with each piece of context provided. So as you provide context to the supervisor as well as each additional sub aent that really helps the supervisor do its job better. The more clear instructions you can give the better it does. Fewshot examples also help here. So if you include clear space descriptions for what your data topics or agents can handle and even few shot examples to teach the supervisor when to use each agent. So give it a couple of examples so it knows how to route. And finally gather feedback from your subject matter experts. This is where you can leverage labeling sessions. If you've seen our review app uh is exactly that right that you can use with other people. And so you can use that with your experts. You can have them label data and again this isn't thousands of examples. You can use just 10 or 12 and you can get market improvements and then that automatically improves the model over time. Now let's jump in and see it. So I'm going to jump back into data bricks here and go back into my agents tab. And this is where this multi- aent supervisor comes in. Just like what Casey showed you during the keynote, you provide a name as well as a description and then you configure up to 20 different agents or tools that your supervisor can use. So I could leverage a genie space, an agent endpoint or an MCP server. I've actually already built this out so that you don't have to watch me type a bunch. So let's look at what that looks like for our telecom example here. I provided a pretty detailed description here because I want the supervisor to know how and where to route questions. So, billing and usage, how much was a bill, how much data was used. Those are some of the examples I want to give it so that it has an understanding of what each sub aent is responsible for and how to route between them. And then here you can see I'm leveraging my knowledge assistant endpoint that I built out where it has my technical support queries and then I have my different genie spaces directly available. So now let's ask it a question. So I'm going to start with one of the ones that we were just using. So um can't make calls but data works. This is a simple question. This is one of the ones that we used when we were testing our knowledge assistant. And so I expect it to route to the technical support agent, which it did right away. So it rephrased the question. Customer reports they can't make calls, but their mobile data is working. So it's now interacting with that knowledge assistant that we just built out, and it's going to give me uh the results of what it finds so that we can do some debugging here. All right. So now it's done its handoff. And so now our knowledge assistant is cooking up an answer for us. So we should see this here in a second. Perfect. So now let's look at what it came together with. So we can see that knowledge assistant searched and found again that full LTE issue that we were seeing, potential SIM card issues, and it provided the footnotes as well. And so that way you can verify your sources and make sure that things are working well. And then our multi- aent went and s and summarized that for me. So, it pulled together potential settings issues or network settings problems. And it even gave me options across both iPhone and Android because I didn't actually specify what type of device I was using here. Now, let's ask a more complex question. So, I'm going to ask something that requires both my structured and unstructured data here. So, I have a customer. They keep seeing speed reduced on their phone. What account plan are they on? How much data have they used this cycle and how can we get their speed back? So, let's see what our supervisor decides to do here. Perfect. So, first it's going to help me check the customer's plan, data usage, and how to restore it. So, it's giving me a summary of what it just heard it wanted me to go do, as well as restating the customer ID here so that I know that it's following the right examples. So, it's going to start by asking the account agent what plan is the customer currently on. So, it's going to do that handoff. So, we can see that they're on a premium individual plan. I don't know what that means. We have lots of plans as a telecom company and so I'm going to need a little bit more context and I suspect that our agent here is too. And so, now it's going to ask now I'll check the data usage for the customer billing cycle. So, it's going to ask our billing agent that Genie space and start to get back their total data usage for this month. And just like me, the supervisor doesn't know all of the details of the plans that we have available. So, is this the same amount of data? Have they exceeded their cap? I actually have no clue. So, it's going to hand that off to our plans and products genie space that we explored. And in here, I can see that I have a data limit of 10 gigabytes. And it also gives me a description of what that plan is. And finally, it's now going to check in with the tech support team to see how to restore the customer's speed. So what we can see in here is that the customer is using a significant amount of data. They're using way more than 10 gigabytes. And so that's very likely the reason. But what are some of those mitigation strategies that we could do? Perfect. So the tech support agent looked across our knowledge base and some of those common FAQs and found that you could purchase additional data, you could upgrade a data plan, you could connect to Wi-Fi just in case you don't realize that you're not on Wi-Fi and using data in that case. And then perfect. So now our multi- aent supervisor has also pulled together a report for me. So I can see that their current plan is premium individual has a 10 gigabyte limit. They've currently used about 463 GB which is a lot more than their current uh plan allows for. And then it gives me a summary of the analysis. So they're seeing the speed reduced because they have used significantly more data than they were supposed to. And so we've given them some options as well to then go and mitigate. This is a really easy way to then take advantage of all of the data that your organization is working on and the various agents that departments are building and pull them together so you can get much deeper and more comprehensive insights. So let's round us out here. You can leverage agent bricks to build data intelligent agents. And so we have knowledge assistant available today. We would love for you to try it. It's in beta. So if you haven't seen this yet, look in your previews tab and you might be you should be able to enable it depending on your region. And then our multi-agent supervisor is coming soon. We don't have an exact date yet, but roughly sometime in the next couple of weeks after summit. So please take a look for that. We really look forward to getting your feedback and for you to see the power of agent bricks. Thank you.