From Prototype to Production with Agent Bricks

Summary

Agent Bricks is the enterprise agent development platform on the Databricks Data and AI platform, providing access to all frontier and custom models via Unity AI Gateway and governance tools that ensure agents respect Unity Catalog permissions without requiring code rewrites when models change.
New capabilities demonstrated in this video include a Supervisor API with a managed harness supporting background execution mode and approval gates, plus agentic tasks that allow agents to be scheduled directly within Databricks Jobs.
Lakebase powers persistent agent memory on the platform, and MLflow tracing with on-behalf-of-user governance ensures every agent action is attributable and auditable end-to-end.

From Prototype to Production with Agent Bricks

The gap between a working agent prototype and a deployed enterprise app is bigger than most teams expect. In this demo-led session, we’ll build a custom agent with the Agent Bricks agent platform and show the pieces required to ship it: multi-model access, governed structured and unstructured data, tools, memory, traces, evals, monitoring, and flexible deployment options. Come learn about the full agent platform on Databricks.
Talk By: Aravind Segu, Software Engineer, Databricks ; Elise Gonzales, Staff Product Manager, Databricks ;

Chapters

FAQs

What is Agent Bricks on Databricks?

Agent Bricks is an enterprise agent development platform built into the Databricks Data and AI platform that provides the tools needed to go from a prototype to a production-ready agent. It offers multi-model access, governed data connections, tools, memory via Lakebase, MLflow tracing, evaluations, and flexible deployment options.

How does Agent Bricks handle model switching without rewriting code?

Unity AI Gateway within Agent Bricks allows developers to swap models—including frontier models and custom models—without rewriting their agent code. This lets teams experiment with newly released models while keeping the rest of the agent architecture stable, addressing a common pain point described in this video.

What is the Supervisor API in Agent Bricks?

The Supervisor API is a managed orchestration harness in Agent Bricks that supports background execution mode and approval gates, allowing agents to run long-running tasks while requiring human sign-off at defined checkpoints. This video demonstrates it using a store operations agent called Coastal Bowls as the example scenario.

How can I schedule agents to run automatically on Databricks?

Agent Bricks supports agentic tasks that integrate with Databricks Jobs, enabling agents to be scheduled on a cron or event-driven basis. This allows teams to automate recurring agent workflows the same way they would schedule a standard data pipeline on the Databricks Data and AI platform.

Full transcript

[00:08] Thank you all for joining us on the last day of data and AI summit. We appreciate you making it through. Hopefully you had a lot of fun at Oracle Park last night. Welcome to from produ prototype to production with agent bricks. My name is
[00:20] Elise. I'm a product manager at data bicks joined by Arvin, one of the engineers on the team. We're really excited to talk to you about some of the developments here and show you how to build agents on agent bricks and with
[00:32] data bricks. Just to give you a sense of what we'll go over today, we'll cover state of the union, some of uh the motivation of why we built some of the new features that we'll introduce. I'll give an overview
[00:44] of agent bricks, our agent development platform. We'll then go through an example and build an agent together. It'll be our store operations agent. So I'll explain what that is, cover some of what's new.
[00:56] Arvin will show you our new supervisor API and agent task and then we'll wrap up. All right, so what motivated us to build some of the new features that we have in
[01:09] agent bricks today? What have we been hearing from our customers? The first one is the new models are really expensive. They're really great, but they're really expensive. Can I try a new model without rewriting everything?
[01:21] How do I experiment a little better? Especially since there seems to be a new model coming out every week, every other week. Second one is how do I make sure my agent respects permissions? I don't want it running a muck. I've invested a
[01:34] lot of time in governance with Unity catalog and so how do I make sure that's respected as I build my agent? And the third one here is my agent needs to understand what different acronyms mean and how my company operates. How do I do
[01:47] that? So how do you add context into your agent? So, as we walk you through Agent Bricks, we'll cover all three of these different aspects. So, introducing Agent Bricks, our
[01:59] enterprise agent developer platform. We had this last year, but we've made a ton of developments over the last year. So, really excited to walk you through what that is. If you were watching the keynote yesterday, this is the introduction
[02:12] slide that Casey showed. So, we have a number of different features within the platform. And so starting here on the left, I'll start with choice. So we're the only platform that gives you access to all Frontier models as well as custom
[02:26] models. Just like we announced yesterday, we now have Kimmy as well as we'll have Grock. And so you can use just any model on the platform. In terms of choice, you also get your choice for orchestration. So you can use
[02:40] hosted OmniGen. We'll talk a little bit about Omnien today. We'll show you it as we uh build an agent together. You can also build custom agents and agent tasks. You can use any popular authoring framework. What I'll show you today is
[02:53] with the OpenAI SDK. We find that one is quite popular. But Langraph, Crew AI, any of these can also be used to build a custom agent on data bricks.
[03:05] The second piece here is context. So, how do you give your agent context about your business, about your goals, and make sure that it can answer with that context in mind instead of hallucinating or inventing responses?
[03:18] So, the first piece here is knowledge, data, and memory. So, we have a bunch of new features that we revealed over day one and day two here. So, support for ingesting and understanding documents, tables, volumes, a new memory API, and
[03:33] Lakebase as a whole powers a lot of our new features here to make sure that your agent understands the task at hand and can remember what their user has asked them before. The other big piece here are a bunch of
[03:47] new tools. So, that's a lot of the investment we've done over the last year is to give you a lot of tools to build agents much more easily. So this starts with some outofthe-box connectors that I'll show you. So we have SharePoint,
[03:59] Google Drive, Atlassian, GitHub, Glean. All of those are available for you to use over MCP. They can be used in Gen1, but they can also be used in your custom agent as you develop on data bricks.
[04:12] The other piece that we've introduced, and we'll show you all of these tools here in a bit, but we have introduced a number of new tools for data bricks assets themselves. So you can query tables, you can query volumes by a tool.
[04:25] Um, and we're looking to make sure all of your UC assets are queryable and usable inside of agents. The final piece here is control. And so under that umbrella, we have deployment.
[04:39] So you get a lot of choice and control over how you want to deploy your agents. We introduced our new sandbox feature yesterday. So you can run code inside of a sandbox, keep it isolated, but still give it access to the data that you have
[04:52] in Unity catalog in a secure way. You also get your choice of agent deployment. So we recently moved and recommended data bricks apps as a really flexible framework. That's what we'll show you here today. But as we saw in
[05:05] the keynote on the first day, you can also use Genie agents to deploy as well. So we're trying to give you some control there. And then durable execution. So as your agent runs, especially if you're doing deep research or longunning tasks,
[05:17] what happens if a computer, if a VM dies, how can it recover from that where you don't have to repay and redo all of the work that it already completed? The final piece here is on governance. I
[05:29] think we talked about this a bunch yesterday, but the Unity AI gateway has had a really big year. It's been really top of mind for folks. So how do you integrate all of that into your custom agents as you develop them on data bricks? So this includes observability,
[05:43] evaluation, authentication, and smart policies. So all of this encompasses what we built over the last years as we've been building out the agent bicks developer platform.
[05:55] So what I'm going to show you here is the store operations agent that we'll be building together. So in this example, Arvin and I are regional managers of a company called Coastal Bowls. It's a Mediterranean
[06:08] restaurant that I invented for this demo. And so what we're looking to do is get some context about our different stores, how they're operating, and be able to give our store operations, regional directors, managers access to
[06:20] that data so that they can ask questions of it. So as we build this out together, what you'll be able to see is that we have a Genie agent in there. So this is our store operations data. So how are they doing against their goals? How many
[06:33] people are working different shifts, information like that. We have two custom MCPs here. So this is in my company, we have our own sort of labor and scheduling MCP which we built and
[06:45] deployed on data bricks apps. So that's a really great option if you do want to connect to your internal systems but maybe you haven't built out an MCP server yet. And then I have a couple of built-in tools here as well. So I'll be using the new code executive tool to run
[06:58] some Python analysis as well as Confluence because that has context for my overall 2026 goals as coastal goals. It'll come builtin with memory. So I'll show you that as well. And then we'll
[07:10] also get a glimpse of some of these platform services like MLflow and Unity AI gateway. So I'm going to jump in here. I'm actually first going to show you. So as we look at the agent and build it out
[07:23] together, I'll have started from one of our app templates. Here we have a number of different examples to help you get started. They all as well come with underclaude. Here we'll be able to see some skills. So this helps teach Claude
[07:36] how to build and deploy your agent on uh agent bricks quite a bit easier. But as we look at this, we have a number of examples across Langraph as well as OpenAI. These are the two most popular frameworks that we see today. But as I
[07:49] mentioned before, you can use any of the popular authoring frameworks that you would like to. So I'm going to jump into data bricks apps here. So this is our new home for apps. If you haven't seen it, it's under
[08:01] our dropdown here. So, I'm looking at the new homepage and I have some quick access over here because I have a number of different agents already deployed. So, we're going to load my agent here. You can see it's
[08:14] running and it'll have some basic tools. So, what we're going to do is I'm actually going to jump in and ask it oh really quickly, what tools do you have?
[08:29] So, we'll see what tools it started out with. I have preceded it with a few of these tools, but we'll also end up adding one of these during this session together. So, we can see it already has the data and analytics genie that I mentioned that has some of my store
[08:41] operations data. It has the two MCPs. So, it has my store labor and scheduling system as well as my ops task. And then it also has a utility to get to current time. That way, if I ask about let's say
[08:53] like the last 12 weeks, it'll actually know what the last 12 weeks means. So, as I mentioned, I started out from one of our app templates. And so, I'm going to jump here into OmniGen. Omnigent is our new meta harness for
[09:06] using coding agents. And so, I'll show you here some of the code that's already available. And so, we're loading our conversation here. Hopefully, it loads here in a second. A little bit slow on
[09:19] the internet. Perfect. So, I mentioned I started out from the OpenAI template. And so what this gives you under my store ops agent here is I have an agents.mmd and cloud.md. That
[09:32] way this helps you understand how to build the agent alongside claude or your favorite coding agent. And then I have a number of uh basically templates already
[09:45] built out for me. So I'm going to jump here into agent.py. And this is where you'll end up seeing the bulk majority of the code you need. And this is where claude and coding agents can really help you modify this because the templates
[09:57] really lend itself really well to being used so you can build and deploy really easily. So when I scroll down here I can see um I can see some of my configuration. So I'm currently using the opus 47 model which really great
[10:10] agent model. I have my genie space. I have my store time MCP as well as my ops MCP. And then I have my whole system prompt of how Coastal Bulls wants these uh questions to be answered. And then if
[10:23] I scroll down a little bit, I will also be able to see. Okay, I do have Lakebase in here as well. So this is something as you set up this template, you'll be able to add lakebase as a memory store. We will as
[10:35] well have memory APIs coming out shortly. So I'll give you a preview of those. And so a lot of this is preceded for me. So all I had to do was give it the tools that I wanted to use. So, let's ask a little bit more complicated
[10:47] of a question here. So, I'm going to ask, show the average versus actual training hours for my to for my stores in my Pacific region and flag any that are falling behind. So, for Coastal
[11:00] Bulls, we have some training goals that we want to meet this year. We want to make sure our employees are staying up to date with everything that they should be doing. So, it's going to first query my Genie space here. This whole template was actually already available for me.
[11:12] So this is something that we preede into all of our templates. So I didn't have to actually build out this UI here, but you can also customize it. And so Arvin will show you that in his demo is some of how he's customized this template.
[11:24] But we wanted to give you a really easy interface API already built out. So I'm going to scroll down so I can see Pacific region training hours last 12 weeks. So I can see I do have a store here that is falling behind. And so it
[11:38] was able to reach out to Genie, run that query for me, and then return and format that response. So then we're going to ask here, why is store 104 behind on training? Is it a scheduling problem or a completion
[11:51] problem? Since we have that custom MCP, I am able to then query the basically scheduling system that I have in here. And so you can see it's going to get store schedule summary. So that way I'm really able to dig in deep with the data
[12:04] that I already have on data bricks and some of the the tools that are already included here. So we'll wait for this to come back. And what's nice is both of these are running in parallel and so I will be able to get answers a little bit faster. So we can see both of them have
[12:17] completed. Hopefully it comes back here in a second. Might have to refresh. We'll see. This is how you know it's a live demo, right?
[12:29] >> Perfect. All right. I think I was just being a little impatient here. All right, so let's look at store 104. So it's both a scheduling problem, but it's primarily a utilization problem. So we can see they've actually been scheduling
[12:42] a whole lot of shifts, but some of them actually got flipped to service. So they were scheduling them, but then they ended up actually needing these people to work the stores. So that's really helpful for me as a operations manager as I'm looking at my region to
[12:55] understand how I may reach out to this person. And so then I'm going to also ask here, what are our training goals for the year? How does 104 stack
[13:09] up? So let's see if it has the context about what my goals are for the year. I don't have a tool that exposes company training goals directly. Okay, so I
[13:21] think we saw this earlier. So, I had Genie. I had my two custom MCPs in the time tool, but I didn't actually have my company context built in here just yet. So, what we're going to do is we're actually going to go back into OmniTent
[13:33] here, and we're going to ask two things. So, one, we've actually been seeing this is running quite well with Opus, but if I am going to release it to all of my store operations managers, regional directors, that could end up being
[13:45] pretty pricey. And so, what I'd like to see here is could we use potentially a cheaper model and still get good results? But let's also ask and see if we can get the Confluence MCP added to my agent here. So I'm going to ask for
[13:58] the agent store ops, can you swap the model to Haiku, add the Atlassian MCP as a tool, and then redeploy the app. So it's going to start working on that for me. And we will see here in a couple minutes when it redeploys the app.
[14:10] That's what's really nice about using data bricks apps as a backend is it's actually really quick to spin up an app and it's also really quick to uh modify and redeploy. So it makes your development cycle much much easier. So
[14:22] while that's running, what I'm going to do is I'm going to close this out. The other thing that we built into the template for you is it does help you with MLflow tracing. And so MLflow tracing here. So I'm back in the data
[14:34] bricks workspace here under experiments and I found my store ops agent demo experiment where I'm storing my traces. And so I can see in here I have a number of the different questions that I have
[14:47] asked. And so I can see that I'm listing tools first. So you want your agent to be aware. This is all actually built into the template for you. So it's not something that you have to think about as you use any of our templates. But you need to know which tools are available.
[14:59] So it's doing that. We can see that it takes a turn. It's generating with uh async data bricks completions. But that just means that it's asking that genie space. So I can see in here it's asking the genie space the very specific
[15:12] question that it's asking here as well. And so all of this comes out of the box with the templates for you. That way you have observability and monitoring already built into your agent. The other thing that's already built in to all of these is on behalf of user O.
[15:26] So as you're building out your app, you can decide if you want to do service principle based off. So is everyone sort of treated the same or do I want to query data on behalf of the user? So I want to make sure that the data that I'm querying is the data I'm allowed to
[15:38] query and should be different for Arbon for example. And so all of that's actually already built in for you as well. So, if I go into my Genie space here, I'm on the Genie spaces tab under SQL. I
[15:51] have my store ops coastal bowls. And what I can do is look here under the monitoring tab and look at the question that I've asked. And we can actually see I was like going to have to zoom out a little bit. Awesome. All of these have
[16:05] actually just been running as me as a user. And so that's really important as you think about governance and as you think about building agents. Whose identity are they using? what data are they allowed to access? And our goal is to make that really easy for you to
[16:18] configure to your specific needs so that you don't have to think about it and you don't worry about getting it wrong, for example. So now let's go back and see where our agent is. Okay, path confusion
[16:30] resolved. So it's still doing some work here. So we'll let it finish that. The next thing that I'll show you here is our Unity AI gateway. So over here under AI and ML, we have our AI gateway. We
[16:42] have a number of MCPS and LLM available to you. So you can see agent integrations, featured models. So the one that I was using is featured here, but I'm actually going to start using a smaller model so I can save a little bit
[16:54] of money for my company. You can see some fastest response and highest traffic. But importantly here we also see MCPS. So Genie, we introduced that on day one. So we do have a genie MCP.
[17:07] We also have Atlassian built in. So this will be the connection that I'm adding to my agent so that I can bring in that confluence context and have a better sense of what that training data is. We also have GitHub out of the box for you.
[17:19] So that way if you want to search code, manage issues and pull requests, you're able to add that directly in there as well as Slack. And again, all of these operate on behalf of your identity. So if I do go into the connection here, I
[17:33] will see that I could revoke credentials in one second. The internet's actually quite slow. All right, I'm not going to do that because I do need my demo to work. I don't want to log in again. Uh, but this is operating on behalf of me.
[17:46] And so all of that again, you don't have to build out the connection to Atlassian. We have that in here for you. It's the same connection that you'll use in Genie 1 as you can use in your custom agents here as well. So now let's go
[17:58] back and see where Claude is. All right, so running the demo redeploy script. Perfect. So, it's made some of my updates for me and I can actually see that within the change view here in
[18:10] Omnien. So, I can see that it's modified my uh YAML file. So, I can see the configuration the app.yaml. So, that's some of the resources that my app needs here. I've been swapped here to haik
[18:22] coup. So, that's really convenient. And then I've also we've updated here the data bricks yl. And so, this is all of the configuration you need and it's built in. And again, we have the skills in here to help you build this out. So,
[18:35] let's see where we're at on our redeploy. Might not be quite finished yet. Give it a second. One hour ago. Okay. So, it's still running this redeploy
[18:47] here. That's okay. So, we'll give it a second to do that. Um, and we'll jump back in. And what I'm actually going to show you here is I'll show you some of our new features and then we'll come back to this redeploy. This is kind of
[18:59] the fun part of using um AI here is you never quite know what it's going to do. So I'm going to show you some of our new features. And so what I've built and used in this template here is actually
[19:11] lakebase as a memory store. So it's absolutely something you can still do. If you want to control your memory, you can. And so that means you can use lakebase and DIY it. But we're also
[19:23] going to have agent memory available via API very shortly. So the beta will launch actually next week and our ultimate goal here is to give you a unified interface for both short and long-term memory. All of this will be
[19:35] stored in UC. So this is a quick code snippet of what will be coming. So you essentially give it a memory store backed by Unity catalog. So you get all of that governance built in. You give it a scope. So in this case the UC entity
[19:50] is actually owned by the app. But when we give it a scope, it knows who I am as a user. So that way I not going to go in and modify my conversations or memories directly in UC. That doesn't totally make sense as a use case. But the agent
[20:03] still knows who I am and is not going to suddenly give Arvin my memories or my conversation store. So we wanted to make that really configurable and easy for you. The next piece here is our data brick
[20:15] sandbox. So this just launched into beta as we announced at the keynote yesterday. So this gives you a really secure sandbox to run particularly code execution within, but it also gives you secure access to Unity catalog data as
[20:29] well as your workspace data. And so you can run it in that isolated VM. Spins up super quickly. Um, and it's all managed on our serverless infrastructure for easy integration and then automatic scale up and scale down. That way you
[20:42] don't have to think about that additional cost there. So let's jump back in and see if our agent has redeployed. Perfect. I think it has. Let's look. I'm going to refresh really quick. One minute ago. Awesome. All
[20:55] right. So, let's launch our agent again here. Let it load. Awesome. Okay. So, let's ask what tools do you have? And you can see here I do
[21:08] have my memory store over on the left. So, this is lakebased. Again, this is something we offer out of the box for you. So we can see we still have Genie, we still have labor and scheduling um utility. Uh oh. If you need to integrate
[21:21] with those, let me know. Okay. Um can you query Confluence for my goal for my training goals? Let's see if that worked. It may just
[21:33] not be sure because it does use on behalf of user off. Okay, that's okay. That part of the demo did not quite work. All right, thanks Claude. We will not debug that live here. It does work though. Uh, and we do have docs that
[21:45] will cover how to configure these out of the box templates here. But I am still getting snappy responses. So the haik coup uh is working quite well here. And so that just shows you how easy it was honestly two lines of code to swap to
[21:57] that new model. And that's one of the benefits of using Unity AI gateway is it does give you a unified API to use models against. And so that way when you do want to swap models even from let's say anthropic to an open AI model,
[22:09] you're not completely having to rewrite your app there. All right. So, I'm now actually going to transition and let Arvin show you some of our new tools to build agents even easier and with even less code than some
[22:21] of the the template that I just showed you. So, thank you.
[22:45] Do you want to cover this? >> Oh, I totally forgot about that slide.
[22:58] have AI parse document and some of our document intelligence functions. Those are now available over MCP. Can obviously query your data as well. So, we have tables as tools. And so you may not always want to build a Genie agent,
[23:10] for example. Maybe you just need one or two tables to attach to your agent. We also have a built-in web search tool now. And so if you do want to add web search, it's really easy to do that. You don't have to go and uh potentially acquire a third party, for example.
[23:23] Dashboards are also available as a tool. And so what does that really mean? What's a dashboard as a tool? Um every dashboard comes built in with a Genie agent or Genie space behind it. So this gives you a really quick way to query
[23:36] that without having to maintain a separate Genie space or Genie agent. And as I mentioned, we have a number of external services available here. So we have SharePoint, Google Drive, Atlassian, GitHub, and Glean out of the box, but you'll continue to see more
[23:49] here as well. And so we'll continue to build these out and again they'll be available for custom agents over MCP also in Genie1. So we want to make this a really composable and easy to use platform. Now take it away, Arvin. Thank
[24:02] you. >> Thanks Elise. Um yeah, for the second part of the demo, I'll demo supervisor API. Supervisor API is our managed harness. Uh this is where we run the agentic loop for you on your behalf so
[24:14] that you don't have to write too much code. Uh we handle everything and it comes with all the bells and whistles such as background mode, streaming, parallel tool calling, uh durable execution and all the features. Um so
[24:27] this is how supervisor API kind of looks like right now. Uh it's a single endpoint hosted under MLflowv1responses. Um you can pick your model, your tools and the instructions which is like your
[24:39] system prompt in order to configure your agent. It is completely compatible with open responses interface. So that is the response interface that most models are capable with. Uh and it is callable completely via OpenAI client uh
[24:52] following the OpenAI standards. It is also model agnostic. uh you can swap between all the models, all the frontier models that we have offering in the Unity AI gateway and it's built with AI gateway in mind. So you get inference tables, rate limits, fallbacks and
[25:05] everything for free. Um and usage tracking is some coming soon on supervisor API. Uh so some of the tools uh that supervisor API comes with uh Elise covered most of them but yeah you get you get your data tools to access
[25:17] your UC tables, vector search, volumes, dashboards, UC functions. Um we also have agents and MCPs. So any agentic serving endpoints, custom MCPS on data bricks apps, custom MCP, custom agents
[25:30] on data bricks apps, knowledge assistance you build with data bricks and super uh agent bricks, genie agents, external MCP servers and also the sandbox web search and memory which is coming soon.
[25:43] So um the other one let let me jump into a demo before we before I cover agentic task. So to start off with I'll show you how simple it is uh to use a supervisor API. This is just Postman. I'm just sent
[25:55] running a query. Um, so you don't need to run write any fancy code. Um, you can just call our AI gateway ML flow v1 responses endpoint. You can pass in a model uh your input which I'm just call
[26:08] I'm just asking how many stores are in our Mountain View region for the coastal bowls genie space that just showed you. You pass in your genace ID, any description to guide the agent in exactly what you want to do and the
[26:20] response comes back in a completely open responses format. you get your answer, you'll see all the function calls, uh, function call outputs, um, and the final answer. So, you can then plug this into any of your existing products just like
[26:33] an API call and you have an agent out of the box. Um, now let me jump into a quick demo. Uh, as Elise mentioned, you can also use databicks apps to customize your UIs. So, I have customized my UI for coastal bolts a little bit using the
[26:46] chat templates. I have three new things here which are back backed by supervisor API which is one I can now generate reports. These are longunning reports which are not mostly a chat interface but something that runs in the background. This is a background mode in
[26:59] supervisor API which runs completely on our compute. So you're not running on your control you're not running on your servers or anything. So you don't have to be responsible for keeping the sessions open any timeout issues making the sessions closed durable execution
[27:13] streaming everything comes built in. We can also have the schedule feature which will allow you to use agentic tasks to schedule this in the future. Uh I'll talk about this a little bit in a bit. Uh and yeah, so before I jump in and
[27:25] show you the code because background mode takes a while, I'm just going to start generating a report. Um and let me also send in a prompt and then we can look at the code together while this works. So coming back into Omnigent um the same repo, the same skills that
[27:38] Alisa showed you, we also have skills for supervisor API. So you can if you want to use this instead you can just ask your cloud to build you something with supervisor API and it will guide you and build you through it. So the supervisor API code is extremely simple
[27:50] uh because we do all the heavy lifting for you. Uh so you start you come in you specify your model which basically uh I think that's better. Yeah. Your tools which is I'm specifying my genie space
[28:02] ID uh any MCP servers that I have and then you would give it a spec. So you would give it all the tools with a description. So my Genie space, my custom MCP server, uh and also my
[28:14] Atlassian connection that I have here. Uh hopefully this works for my demo.
[28:28] where are my helpers? Helpers. Uh sorry. So yeah, this is how we're making a client. Um, so if you guys are familiar
[28:41] with data bricks apps, the obo token which is the which is the current users token is passed through the headers to the x forwarded access token. I'm taking that in and I'm initializing data bricks openai and I'm basically saying use that
[28:54] token for my authentication and I'm saying use AI gateway equals true and use AI gateway equals true basically toggles the AI gateway supervisor API functionality and now you're using AI gateway supervisor API under the hood.
[29:06] So now coming back in here uh we create that client and then we just use the open AI terminology of client.responses.create create uh pass in the model, the instructions, the tools. I'm enabling streaming here. Uh
[29:20] for this demo, I'm saying parallel tool calls false. Uh you can set that to true. And I'm also passing in some extra headers. And this is really important for us for tracing. So if you want distributed tracing in your services where you have clients and services and you want to tie all those traces
[29:33] together, you would pass in the parent trace. So like the trace that comes from here goes into the headers and then I pass in the trace destination in the extra body. Um and that's it. And you have an agent up and running. So now if
[29:45] we go back to our demo, um, there we go. So this is running. Oh. Oh no. Tool call error. >> Cloud is not on our side today apparently.
[29:57] >> Let me retry. Okay. Oh, there we go. It's starting to pull some region.
[30:10] Let me Perfect. So, as we noticed, uh, he set parallel tool calling to false. So, it is kicking off a genie query and then it'll kick off another one. So, we'll see that pop
[30:22] up here in a second. But, as you saw, the supervisor API is significantly shorter than even some of the the templates we provide for you for some of the other frameworks. So, it can be a really great option if you don't want to invest in a ton of customization and you
[30:35] want the built-in tools available for you. >> Yep. So, it is going off. So now it says query genie got some information uh query confluence um it got my FY26
[30:47] training page it got all that information uh from Confluence MCP server um it's now getting the confluence and then it's also we also have approvals built in. So if we have if you ask custom MCP servers we have
[30:59] approvals built into supervisor API and our templates so they surface in the UI. So now it's telling me I'm going to call the get store schedule summary from our custom MCP with these parameters. Do you allow or deny? So governance is built into supervisor API and once you click
[31:12] allow the the agent continues um and produces the final result in a bit
[31:26] >> being very thorough >> being very thorough here while that's happening um I just noticed that our background agent is also done uh so I'll just open the report here so this is basically something I've kicked off in the background uh in like to show you how we did that in code I'll go back
[31:39] here. Um I'll go into my supervisor API background mode and the only difference that I've sent here is background equals true. Um there's no other difference. Uh once you once that happens, it
[31:52] immediately returns the response and the compute and everything is happening in the back end and we'll just let you know via streaming or polling when that stuff is done. Uh so now if you come back here uh you can look at the report that we
[32:05] have generated. Uh this is the report that's showing all the stores in our coastal bowls, what the training percentages are, what the food safety is, the labor laws, and the risks of if it's a good store or a bad store, and if you as operation manager need some
[32:17] attention. We also have tracing built in. So if you have background mode and you want to know how to arrive to these conclusions, you can go into our traces. Um and then this one's the first one. Uh so complete detailed traces of how it
[32:30] like same thing calls the LLM calls genie spaces and walks through the steps in order to get all the information that it needs. Um going back to our chat
[32:43] looks like it's done. Um it's basically saying it's a completion problem not a scheduling problem. The evidence was over the last four weeks sore 10007 had no training shifts which we saw earlier with Elisa's demos as well. Um and then
[32:56] five out of eight scheduled training shifts were actually completed. Uh same thing here uh traces are all uploaded to UC. So the conversations uh here are in this experiment. Uh so you can go in
[33:08] here and look at all the traces that um you have uh that the agent has called out to to answer this question. And the final thing I wanted to show you was uh schedule feature. This basically allows
[33:20] you to schedule this agent uh on a on a trigger or a schedule basis. uh using our new agentic task on data bricks jobs. Uh I've already gone and ahead and done it. So I'll just show you the UI
[33:32] here. So when you come into uh oops there we go. When you come into uh data bricks jobs now when you click add a task with all the tasks you have a new
[33:44] task called agent. This is now backed by supervisor API for you to create an agentic task and you get all the bells and whistles of jobs. So you can schedule them, you can add triggers, you can make your own DAG. So if you have a complex workflow and want to kick off an
[33:56] agent in the middle to like act on some of some of the stuff and then make some more decisions, everything is built in. And it's everything is configurable via the UI. You get very similar interface to supervisory API. You would say your input, select your model to any of the
[34:10] models that we offer at in Unity AI gateway, any system prompts such as instructions you want to give the model and the tools. And these are all the tools that we have talked about and they're available. So we have built-in tools, genie spaces, search indices, UC connections and so on and so forth. And
[34:24] you also can configure trace destinations which basically tell you where the traces should live and you can schedule them. So I've been running a couple for the past uh every 3 hours uh for the demo and you can just click into it uh see what it did and then I've been
[34:38] generating reports. So this is all the stuff that it has done and that's the report has generated. Uh so you can use this to orchestrate your jobs and workflows as well. Um with that I'll just go back to my slides
[34:51] which are here. Um so agentic task uh it's going to be coming in beta very soon. Uh and use a supervisor API to integrate models tools and traces and I'll hand it out back to Elise for a
[35:04] quick conclusion. >> Perfect. Thanks Arvin. Awesome. So to round us out here as I covered agent bricks is our agent developer platform. It gives you choice, context, and control. All of this is
[35:17] built in for you. As we saw, supervisor API connects directly to Unity AI gateway. So you can get MCPs, models, still have auditability, traceability, lineage, and control.
[35:33] Last thing we'll leave you off with, we'd love for you to deploy an agent today. This leads you to our data bricks apps template repo. And so we would love to see the agents that you build. So thank you so much for all the time. We're available. We'll be available
[35:45] outside the room if you have any questions for us. So, thank you very much.

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.