Building Higher Quality Genie Agents (fka Genie spaces): How to Ensure Accuracy and Trust

Summary

This video explains how to design and scale high-quality Genie Agents—formerly known as Genie Spaces—using a five-step framework: define purpose, curate data, build test sets, add context, and monitor in production.
Unity Catalog Business Semantics (Metric Views) reduce query ambiguity, while Genie Code and knowledge mining automate the initial setup of Genie Agents for faster bootstrapping from scratch.
Genie Agents are evolving beyond text-to-SQL, with upcoming capabilities including file and document Q&A, MCP tool connections to external applications, and scheduled agentic task execution.

Building Higher Quality Genie Agents (fka Genie spaces): How to Ensure Accuracy and Trust

High-quality Genie Agents (formerly known as Genie Spaces) are built on a foundation of clear metadata, curated semantics, and continuous iteration. New capabilities make it faster than ever to bootstrap and develop Genie Agents. In this video, we'll show how to design, refine, and scale Genie Agents that consistently deliver accurate, trusted results. Topics include: Using Unity Catalog Business Semantics (Metric Views) improve accuracy and consistency from the start Leveraging Genie Code and knowledge mining to automate space setup How to define metadata and context in the Genie knowledge store How to use evaluation and Benchmarks to measure accuracy and guide improvements Best practices for monitoring and optimizing Genie Agents at scale
Talk By: Hanlin Sun, Product Manager, Databricks ; Shah Amini, Engineering Leader, Databricks ;

Chapters

FAQs

What is the difference between Genie Agents and Genie Spaces?

Genie Spaces were primarily text-to-SQL tools that answered structured data questions, while Genie Agents are a more capable evolution that can go beyond tables to answer questions on unstructured files, connect to external apps via MCP, and take actions in those systems. The transition from Genie Spaces to Genie Agents is driven by customer feedback requesting capabilities beyond what Spaces could deliver.

What is the five-step framework for building high-quality Genie Agents?

The five-step framework covers defining the agent's purpose, curating the right data set, building a test set to measure accuracy, adding context through the Genie knowledge store, and monitoring the agent in production by analyzing real user questions. Each step is designed to reduce ambiguity and ensure the agent delivers consistent, trusted results.

How do Metric Views improve the accuracy of Genie Agents?

Unity Catalog Business Semantics—specifically Metric Views—provide clearly defined, unambiguous business metrics that Genie Agents can reference when answering questions. Without focused data sets and metric definitions, conflicting definitions and overloaded instructions are common pitfalls that lead to inaccurate agent responses.

How can teams deploy Genie Agents across development, staging, and production environments?

Genie Agents can be promoted across dev, staging, and production environments using the export and import API, enabling a reliable CI/CD path for agent deployment. This video also covers using benchmarks to measure accuracy at each stage and monitoring production agents by analyzing the questions users are actually asking.

Full transcript

[00:08] Welcome everyone. Really appreciate you all sticking with us to the very end of data and AI summit. I'm sure you guys are very excited to hear more about Genie. Um and really excited to be talking about how to build quality Genie agents. My name is Hanlin. I'm a product
[00:21] manager here at Data Bricks and I'm joined by Sha, our engineering manager. Yeah, let's go ahead and start with a quick agenda. We'll start with the level set on what exactly Genie agents are. Next, we'll cover the best practices on
[00:33] how to curate your Genie agents. And from there, we'll discuss how to effectively deploy and monitor your agents in production. From there, Shaw will tie it all together with a live demo showing you folks how to build a Genie agent from nothing. And at the
[00:46] end, we'll definitely make sure to leave plenty of time for a Q&A session. So, I'm sure you've all heard a lot about Genie this week. And I'd like to first level set on exactly what Genie agents are just to make sure we're on the same page. So on the data bricks
[00:59] platform, Genie is our suite of AI features that helps your users accomplish any of their tasks. There are three main experiences. First is Genie 1. This is the simplified home for your business users to talk, interact, and
[01:12] discover their data all at one place in one entry point. Next, Genie Code is the AI coding and data assistant for your technical practitioners. This helps people in the workspace do things like build ETL pipelines, set up warehouses,
[01:26] manage compute, etc. And finally, Genie agents are the low code agents that your users can build to tackle any specific business area. So, as we talk about Genie agents today, I wanted to first acknowledge that Genie
[01:39] agents are actually really just an evolution of what users are leveraging today, which is Genie Spaces. Up until now, many of you have built and deployed Genie Spaces as a texttosql tool. Now,
[01:51] they've been super helpful in answering your users data questions, but users hit a wall when they want to go beyond tables. We've heard from our customers that they want to go beyond what Genie spaces can do today, but had to turn to other tools to accomplish this. Your
[02:05] feedback has led us to evolve Genie agents past Genie Spaces that are fully capable of all the tasks that your teams need. So very concretely, your Genie agents will soon be able to answer questions on unstructured files and
[02:18] documents and also connect to external apps via MCP to retrieve external context and even take actions in them. So just to make sure I I'm I'm clear and I reiterate, you're not going to have to enable another product or your you know
[02:30] your Genie spaces are going to be deprecated in any way. Those products you're already using today will simply just have these new capabilities. It's been amazing to see how customers have been leveraging Genie agents. Meta
[02:42] has built spend and budget Genie agents to deliver fast, accurate insights across their finance organization. General Motors has built over a thousand of their Genie agents in production, completely transforming their data democratization. Insights that used to
[02:55] take hours to dig out can now happen in just moments. So let's take a look at how customers like GM and Meta have been able to build these effective Genie agents. So first I want to give a quick big
[03:08] picture of the overall process. It starts with defining a very clear purpose for your agent. Genie agents are meant to help with a specific topic like product usage or finance, not everything
[03:20] all at once. This focus is what enables higher accuracy. Next, ensure that your data is well prepared. The goal is to really provide a focused well doumented data set so the agent is able to effectively retrieve information
[03:33] based on your user prompts. Next, after you've set up that clear purpose and underlying data, go ahead and start testing questions to your agent. And as you identify potential knowledge gaps, you can fill that in with context. To help you do this at
[03:47] scale, we also have benchmark so you can run a curated set of common questions that your users will be asking. And you can also leverage Genie Code as a assistant to help you build your agent. Deploying your agent isn't the final
[04:00] finish line. It's really important that you continue monitoring your agent in production and refine its performance over time. It's really helpful to think about a Genie agent almost as like a mini employee, right? A dashboard maybe you deploy once and it kind of lives as
[04:13] a static product. But a genie agent is really an employee that as it takes on more tasks and questions from other users, you can enhance it, improve its performance, and take on even more different tasks.
[04:26] So, let's walk through each of these steps in a little bit more detail. After you define a specific purpose like your finance bot or an HR agent, the first thing to do is really prepare well- modeled data sets for it. We highly
[04:39] recommend using metric views as this allows you to centrally define and manage your business metrics in Unity catalog. So this way I can define a central way to calculate sales and apply that across all of my Genie agents in
[04:51] production. It's really important that you provide a focused unambiguous data set because when you have too many fields and overlapping definitions, this can confuse an agent and it's also harder for your authors to maintain. The
[05:04] parallel I always like to keep in my mind is again genie agent is a new starting data analyst at your company. If you give a new starting data analyst at your company two different ways to calculate the same metric, they're going to be confused and it's not going to be
[05:16] consistent answers. Genie agents has the same kind of concept. As a best practice, you can also edit a lot of this metadata on your tables locally inside of the Genie agent. So, for example, if your tables and views
[05:28] and Unity catalog lack those descriptions and you may not have edit permissions on those gold standard tables, you can actually go ahead and just add those descriptions and table relationships directly inside of the Genie agent to model it more
[05:40] effectively. When we introduce files into Genie agents as well, there's a set of best practices that really prepare it for better accuracy. First, make sure you minimize overlapping files such as
[05:52] different versions of the same content, as this can confuse the agent which is the right file to leverage. Next, make sure you add digital native PDFs over scan PDFs as they provide more selectable text that the agent can
[06:05] reason over to answer your specific user questions. And finally, add clear purposes for each of the volumes, which are basically folders for your files. So for example, if I have, you know, a sales analysis genie agent and I want to
[06:19] bring in both win- loss reports as well as, you know, initial pipeline creation descriptions, I want to make sure that those folders are clearly documented and explain what each purpose is for. So the agent knows when to go to each volume.
[06:33] So now that you've defined a specific agent and you've prepared your underlying metadata, you can now go ahead and test those questions and fill in the context. A couple best practices as you detect gaps in your agents
[06:45] ability to answer questions is to first make sure you prioritize those table and volume descriptions. This uplevels your agents baseline understanding of what data is available so it's able to effectively leverage it across all of your different questions that come in.
[06:58] Next, make sure to define SQL expressions and examples. This teaches your Genie unique business semantics and they're also applied dynamically per user question. Additionally, if you add text instructions, which are general behavior
[07:11] guidances for the agent, make sure to phrase them as explicit rules. So, if you give something really general to the agent, like, you know what, uh, push back on ambiguous questions from the user. It's really hard for the agent to exactly understand what that means and
[07:24] when to apply that. So, when you give it very explicit rules, like when the user asks for sales, make sure they include a time range and a quarter to slice this by. then it's a lot easier for the Genie agent to achieve that clarification
[07:36] behavior that you're looking for. And finally, as an author, it's really helpful to leverage Genie code as an agent to kind of help you set up your agents. We'll walk through how exactly this works in a couple slides from now.
[07:48] Some common pitfalls to avoid. First, don't overload your text instructions as they're a lot harder for your authors to manage over time and they could potentially lead to conflicting instructions. That's the second pitfall to avoid which is again avoiding telling
[08:01] the agent two different things to calculate the same metric. Additionally, when we enable file support, please avoid adding large files that are over 50 megabytes as Genie won't be able to parse that completely
[08:14] and analyze those documents. It's a lot more effective to break them up into separate files. And finally, don't add SQL logic as text. Prioritize those example SQL and SQL expression functions.
[08:26] As you're evaluating more of these questions, we know that it's hard to manually test every single prompt. And so we've built the benchmarks feature to help you programmatically evaluate these agents effectively.
[08:38] So here I can define a set of benchmark questions and the gold standard answers that should make it correct, as well as even an LLM as a judge criteria. to support benchmarks. We've also launched agent mode so that you can
[08:51] evaluate these complex user prompts and run them in benchmarks in agent mode. We've also launched the benchmarks via API feature so you can programmatically execute these benchmarks over time and even run them
[09:04] on a schedule to help authors improve their space their agents as they run these benchmarks. We actually suggest context improvements. So based on what we see from what Genie outputs as well as what the gold standard truth is, we
[09:17] can try to analyze what the difference is as well as take into account what context you already have. From there, we can suggest improvements so that you can actually see, hey, like I'm missing a join relationship between these tables. That's why I'm missing all of these
[09:29] different benchmarks. So we talked a little bit before about how I can use an agent to help build
[09:41] specific steps in the process. First is the initial creation of an agent. Instead of having to manually select all the relevant tables and manually add a bunch of instructions, I can actually just go into Genie Code and describe my
[09:53] agents purpose, some of the desired questions they want to answer, as well as the required data. From there, Genie Code will actually go into Unity catalog, identify the relevant context and metadata, and then build your agent.
[10:06] So it'll drop in the right tables, hide unnecessary columns, add missing join relationships, definitions, etc. to give you a really high quality baseline agent to start testing. Additionally, if you're looking to
[10:18] centrally model your fact-based information, you can use Geniode to set up metric views as well. Genie code can also help you debug answers as you're testing your agent.
[10:30] So, you know, as an example, you know, I may ask a question to Genie and it misses a specific filter relationship. Instead of having to manually go through and see what context is missing to apply that, I can just go ahead and describe
[10:42] to Genie Code what the exact issue was and ask for help. From there, Genie Code will debug your desired behavior by looking at the available context and the issue provided. It can then suggest
[10:55] actual context improvements that you can accept. And we always make sure that you as the author have final say over what improvements are made. This also works across your benchmarks. So if you have again a large set of benchmark results and there's a lot of different cases
[11:08] that you're trying to improve on, Genie Code can analyze all of that at once to suggest more targeted improvements. So now that you've created your agent and prepared this context, let's walk through how to deploy and monitor it in
[11:20] production. First, you have a number of options on where to deploy your agents. For we have Genie 1 and Genie mobile which are our native business user homes and we also have native integrations into thirdparty applications so you can reach your users
[11:33] where they work today. You can deploy Genie agents into teams, Slack and M365. And finally, if you want to build custom applications or agents, we also have APIs, iframes and MCPs to really bring
[11:46] Genie to where your users work. We also have a lot of ways for you to monitor your agent in production as well. Every Genie agent comes with a monitoring tab that lists every single user message and their feedback so that
[11:58] authors can review that effectively. We have list conversation APIs as well. So if you want to programmatically access that information and centrally put that into a table and you see you have that option.
[12:10] Genie code also helps with this monitoring step. So, we recently launched this new feature where you can click the analyze space usage button. And what this will do is actually go through all of the recent messages that have been asked and the feedback given
[12:22] by users to give authors a summary. You can see the top questions people are asking, what are the common issues reported and even suggested improvements. So, you can really have this flywheel of deploying your Genie to production, getting more and more users
[12:35] onboarded, and then refining the context as they ask more questions. So that's a quick slide walkthrough of how to build a Genie agent from scratch all the way to man monitoring it in production. I'll now hand it over to
[12:47] Shaw to show you this in action. >> Thank you Hanland. Hey folks, I'm Sha Amini. I'm the engineering manager for Genie agents and also chat in Genie 1. For today's demo,
[12:59] uh the scenario is that I'm a sales manager at this company named Acmeico. And I have a table of opportunities and uh for sale and some of these I have won, some of these I have lost and I
[13:13] want to analyze it and better understand um these opportunities. So, first I'm going to start Genie Code.
[13:27] >> The Wi-Fi >> if the if the Wi-Fi allows me to. Uhhuh. >> I think it's the
[13:40] the global connect. Okay, that's better. Some I think the global connect you might disable.
[13:54] >> Yeah, because you don't have that for our workspace. >> Okay, let's try that again.
[14:06] >> Oh no, the Wi-Fi. This is also the reason that I carry my phone with me. So, we'll see if my we'll have a better better chance with my phone.
[14:22] There it is. >> All right. There we go. So, when in doubt, carry your phone and and tether. Uh, so did I mention that as an engineering manager, I always plan for redundancy?
[14:35] So, all right. So, I'm going to open a new tab with Genie code and I'm going to ask it to create a Genie agent
[14:48] to analyze my opportunities table in the Shaw Amini schema. Add any relevant tables.
[15:09] All right. So here, Genie is actually sharing it thinking traces with you. It's going through the schema, looking at the different tables that I have. It found the opportunities table and it's telling me that it created a new opportunities analysis uh genie agent.
[15:21] So I'm going to go ahead and uh accept that and I'm going to open this space. All right, here we are. So when I go into the configure tab, I can look at the
[15:34] different tables that Genieode has actually added to my Genie agent. I click into these tables. I see that for some of these, I don't have a description. Some of the columns are lacking in terms of the column comments.
[15:48] And so I can ask Genie Code to actually recommend descriptions and and comments for me. This provides additional context that will then enable the LLM to get more accurate insights. So, I'm going to
[16:01] go here and tell it to uh look at the table descriptions and column comments.
[16:23] All right. So, Genieode is now doing exactly that. It's looking at each of these tables. Once it reads them, it tries to figure out if there's a description or a column missing
[16:42] It makes a number of recommendations. I'm not going to try on the auto approve here just because I want to share the approval process for each of these. You can review what Genie Code is recommending and then approve each one to make sure that you really understand
[16:55] what are the different stages to updating this table. So, I'm going to ask it to apply all the suggested descriptions. It's now going to ask me if I want to approve.
[17:16] just to add some additional color as this is going. It's able to generate these descriptions from looking at the context in Unity catalog. So it's not just looking at the table name and sort of making up its own description. It's saying, hey, what's the sample data available? What are the queries that Shaw have written in the past that lead
[17:29] to its definitions that it suggests? Yep. And so now when I click on the tables, you'll see that there's a column description and also comments as well. All right. So now that I have these,
[17:43] there are other things that you could do. For example, you could uh hide columns and so on. I actually want to take a look at the tables and see if I can have any joins uh measures, fields, and filters added. So I'm going to ask
[17:55] Genie Code for exactly that. Uh can you suggest some joins, filters, measures, and fields?
[18:10] And similarly, it's going through and looking at the foreign key and and primary key relationships as well as interesting insights that it could find with my uh with my tables and provide suggestions for those. So, I'm going to
[18:23] go through these. Uh I see that the joins make sense for me. It's mapping account ID to account ID. Some of my uh other tables have an owner ID mapped to a user ID. That's actually correct. And so what you want to do here is just make
[18:37] sure that the information makes sense to you. And I'm going to ask it to add all of these. So again, it's going to review and ask if it could add those knowledge snippets for me. If you were in our demo last year, you
[18:51] would see that a lot of these steps are kind of you point and click and update things with uh your mouse. Now, Genie Code is doing majority of that stuff for you. So again uh you can see the joins
[19:03] added here under SQL expressions. I have filters, measures and and fields added as well. So this is a good place to be. I want to go and ask a question to this table now. So I'll ask looking
[19:18] at the past four months of data what account names have open opportunities
[19:34] and now Genie my Genie agent is actually using all the context that I've added to this uh Genie agent through the data and the semantics to give me the results. Once it has a plan, I can actually go through and uh look at the queries that
[19:47] it's executing. So here I can go ahead. This is an agentic process. So it's actually looking at uh opportunities that uh were not enclosed. So they were
[19:59] still open opportunities and it's looking at the past four months of data. The first uh query actually is not returning any results. This is how you know it's agentic. So it will look at the queries and continue improving and updating his it queries and it is
[20:12] getting more results. Uh so finally it's going to give us the name of the accounts there 105 unique opportunities and I have the list of account names here and it provides me the sources that
[20:26] it has used. All right. So let's say I have now prepared my Genie agent and I'm going to start moving it in production. How do I actually see how people are using this
[20:38] space? So, this is the Acme sales opportunity agent. Hanland and I have collaborated in putting this together. I can look at the monitoring tab. And inside the monitoring tab, I can
[20:51] actually see different conversations that uh have happened with this agent. And from there, it actually shows me a summary of these different conversations. So, uh, total number of questions and how those have changed in
[21:03] the past week, number of unique users, thumbs up and thumbs down. I actually would want to maybe read through all these conversations, but there are actually a lot of them. So, to help you scale and monitoring this stuff, we
[21:16] actually leverage Genie code to analyze the Genie space for you. So, now it's looking at the monitoring metrics and the recent messages and it's coming up with the list of common questions that people ask. uh questions where there
[21:30] might not be as much uh clarity or issues and then it also makes a set of recommendations in terms of how to improve your space. So it's going through
[21:45] so in one way it's good that we lost our connection because now you know all this stuff is happening live. All right. So, I have my top question topics and going below this, it's going to give
[21:58] us general uh pitfalls or places where there are issues or common issues. For example, some people ask for a specific acronym that is not already defined within my genie space. So, I
[22:11] could go ahead and add that here. It actually gives me a suggestion for what that acronym is, which is the demo pipeline expected value. If I wanted to, I could go through these and ask Genie
[22:23] to add that context back into my page. So, next up, we have benchmarks. For those of you coming from a software development background, these are similar to unit tests for your Genie
[22:37] agent. What you'll see in terms of the difference since our last demo is that now you can define benchmarks both for agent and chat mode and these basically include the question
[22:49] that you would ask Genie the ground truth SQL and now you also have an evaluation note that you could add so for example maybe the ground truth SQL is actually a subset of what I would expect Genie to return or maybe it's a
[23:02] superset and I can provide an evaluation note to say exactly that for when I'm evaluating the benchmarks. So, I can run the benchmarks here. It will take me to the evaluation tab and start kicking
[23:14] those off. So, you'll see this going through. We don't have time to do a whole run of these, but I can actually go to the evaluations and see what those uh benchmarks uh were doing in the past. As
[23:27] I'm going through and changing the context and instructions to my Genie agent, I want to run these benchmarks to make sure that I'm maintaining a high quality bar for my uh Genie agent. And also I can now leverage APIs to make
[23:40] these runs. So maybe I set up some sort of recurring job to run the benchmarks for me. All right. So now that I have this uh are other ways for me to
[23:52] productionize it. So a few is that you can use our APIs to export and import Genie spaces. For example, if you have different workspaces for devs, staging and prod. You could use those to take
[24:04] your Genie agent to the next level. Uh for example, in production, we also now have DABs or the declarative asset bundles uh that you can use to also productionize your agents. So what I
[24:17] want to do here is actually embed this agent within another page. So, I'm going to take the URL that I have here and I'm going to go to this wiki page. Add a new page.
[24:36] All right. Let's see. Oh, this one might actually require the VPN. >> Yeah. >> All right. Let's see if I can also connect my VPN over the phone.
[24:50] Impressive. It worked. Okay. So, this is Acme Genie agent.
[25:04] I'm going to add an iframe here. Here's my URL. I copied an extra quotation. Under the width, I want to put 100% and maybe something like 800 pixels is something that I'm happy with. All
[25:18] right. Did that 800 pixel actually land?
[25:43] All right, so here's my Genie agent. Let's see if it will actually load for me.
[25:56] Okay, I hope you get the idea. The Genie agent actually does get embedded in this space. We'll come back to it and see. I don't control confluence, unfortunately. All right, so um next up
[26:13] I wanted to show you. Okay, let's see. I guess this is this is coming out. Next up, I wanted to actually show you uh where we're going to take Genie agents next. So, we heard from a lot of you that you also want to be able to analyze both unstructured data as well as a
[26:26] structured data. And we are positioning Genie agents next uh to be able to be more of a general agent. And so here under my data tab, I've actually added a
[26:39] volume of different uh PDF files. So, as a sales manager, it would be common for me to have reports for win and losses during the opportunities. And so, one thing I'm going to do is ask Genie a
[26:53] question about, uh, losses in the education industry. So, uh, give me a breakdown of
[27:06] losses in the education industry. Here what Genie will do is actually go through and run queries based on the the structured data that I have in my Genie space.
[27:27] All right. So it's going through the thinking traces. I can go ahead and click into those. It now has queries for each of the different uh segments in the education
[27:42] industry. I can look at the code uh see that the industry is actually set properly here. And it's now trying different uh different queries for getting the breakdowns. And here maybe
[27:55] once uh I'm satisfied with the answer, I can always click on uh getting to a result faster. But uh here it actually ran a few queries and it tells me that actually across the education industry
[28:08] uh on the enterprise side is where I have my most number of loss opportunities. All right. So what I want to do next here is actually look at some of those win- loss reports. So looking
[28:20] at the win loss reports,
[28:35] So it's now going to look in that volume where I have my PDF files. Hanlin already shared some best practices here. Uh basically you want to make sure that your PDF files are curated just like your data sets for the right domain and
[28:47] for the type of questions that you want to be asking. True PDF files. So things that are u native digital PDF files that you can select are uh more performant
[28:59] than a scanned ones or ones that have large images. So if you have a choice between a true PDF or something that's scanned, pick the one that's more easily machine readable. I can click on the
[29:11] thinking traces now and it's looking through the win loss reports. So we'll give that a minute to run.
[29:30] All right, it's looking at the different loss rates and now looking at the volume as well. So what you can see here as well something I want to additionally highlight is like this level of insight across different data sources is also
[29:43] what's possible now. So many of you may have built you know a se separate vector search endpoint to answer questions on documents and then try to wire that through you know a supervisor agent framework alongside your tables. Now one
[29:55] agent can reason across both and what we see is that that level of granularity on the tool control actually bodess for better accuracy in final reports because it's not kind of a summarized an insight on the documents being submitted to a
[30:08] supervisor agent. It's kind of one agent being able to leverage both as tools. All right, it's looking at the volumes now. And voila, we have our answer. Uh, so
[30:22] this is context that's coming from the unstructured data or the PDF files is telling me that one of the competitors, Brightwell, uh, is actually winning more of the offerings by just offering AI
[30:34] generated notes as more of a st uh, table stakes feature. The product that Acme is building actually has many other uh, features. for example, for cost control or or admin, but because of
[30:47] that, it's actually more expensive. Uh where the competitor is beating Acme is uh basically by cutting on on prices. So, it's actually being more uh efficient in terms of prices for the features that people want. And so, it
[31:00] gives me a good summary of uh places where I could make some changes. So, with that, I'm going to hand over to Hanlin for going over our road map. Thank you, Hanlin. Awesome.
[31:13] Cool. So, as you saw, we were able to really build effective agents a lot more faster than before by actually leaning on other agents to help you set those up. And it's not just initial creation either. It's debugging, it's monitoring,
[31:25] it's the whole process. And we're really excited to be launching some net new capabilities as well. So, the experience that you saw with the unstructured files, we're looking to launch a first version of that uh this month.
[31:37] So, you know, document analysis is one of the big exciting launches and we'll soon also be launching thirdparty tool connections. So, if you have data inside of SharePoint, Google Drive, these other systems of record, we want you to be
[31:49] able to analyze alongside your structured tables as well.
[32:02] It's not you have to Okay. Um, but these two features are also part of a wider road map theme, right? There's always three things we're trying to do inside of Genie agents. The first is enabling net new capabilities. So in addition to files and thirdparty
[32:14] MCP connections, we're also looking to allow you to define custom MCP tools and plugging those into Genie agents as well as scheduled task and workflow. So maybe there's a common analysis that you want to launch to Genie agent. You should be able to just schedule that and let it
[32:26] run over time. Next, we always want to make sure we're improving the agent development toolkit as well. So as authors and creators of agents, you have everything you need to build them effectively. So as you saw, we are going to continue improving that Genie code
[32:38] for Genie agents experience. We're also going to be launching expanded benchmarks so you can test MCP tool calls and files as well. Uh and also making sure that all your CI/CD requirements are checked off with Terraform and Git folders. And as Sean
[32:51] mentioned, we just launched the the data bricks asset bundle support. And finally, we always want to make sure that the non-technical users, the real final audience of these agents are able to leverage them effectively. So to help these business users, we're first
[33:04] launching chat sharing so that they can actually circulate those insights with other users and even share them to data professionals to double check the SQL and the insights. We're also looking to launch clarification questions. So, you know, if a user inevitably asks a pretty
[33:17] ambiguous question that's not really related to the data, we want to be more push give more push back and let the agent actually kind of walk the user through exactly what they're looking for. And also with access, right? So, making sure that when they access Genie
[33:30] agents via API, they get the visualizations, they get agent mode, uh, and also easier access with app delegated authentication. So that wraps up the the kind of slide and demo portion of this session and
[33:43] we'd love definitely love for you folks to you know ask us some questions and and we'll definitely be able to stick around after the session as well.

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.