Agentic Apps: How Big Companies Are Actually Deploying AI Agents

0:00 I focused a lot on building personal agents on my channel this year, like your second brain, coding assistants, even customer support agents. They're all the same where it's just a single

0:09 agent you're building a use case around. You have that conversation in some kind of chat input where you give it requests and you read through the answers that it gives back. And these kinds of agents

0:19 are fantastic. They run a lot of my business. But when I actually go into a company to integrate agentic AI, a lot of the times they need something more holistic. Now, especially because of the

0:29 AI craze the last couple of years, it is very tempting for people building agents to make the agent the entire automation or application. Instead of it just being a tool within a larger, more

0:40 deterministic workflow or application, which is actually what most people really need. They need what is called an agentic app. This is Oracle's official definition and I love it. Basically with

0:51 aentic apps instead of thinking about the agent first like what are we going to automate with this AI agent we start by thinking about the business outcome right like how are we going to help

1:00 someone with part of their role building the experience for that and then once we have this mapped out thinking about how we can build even multiple AI agents into the application to help with some

1:10 of the automations and honestly this is just the most practical way to integrate Agentic AI if you focus on the business outcome first that means you're much more likely to actually deliver value to

1:20 the organization instead of just integrating agents for the sake of getting on the latest technology, which trust me, I see that way too often. And when your AI agents are simply a tool as

1:30 a part of a larger app, it gives you a lot more control, right? You're building more determinism because you're creating these workflows that use LLM throughout the application, but it's not the agent

1:39 driving the entire thing all the time. So, I hope that makes sense why we care about building these agentic apps focused on a business outcome. And so in this video I want to get very practical

1:49 with you and show you what this looks like to go from a chat application like this like this is my second brain to an agentic app. We have many agents working under the hood to help us with a

2:00 business outcome right like this is helping us manage a role that we have in the business. Now this is more enterprise focused but even if you're building personal applications there's a

2:08 lot that you can learn from this. There's also a lot of different ways we can build these agentic apps. Now, out of all the platforms I've seen out there, Oracle AI Agent Studio does it

2:17 the best. So, we're going to be showcasing Oracle AI Agent Studio throughout this, and I've also partnered with them to bring this video to you. But just in general, there's so much we

2:26 can learn from how Oracle AI Agent Studio builds these agentic apps and all the important things that go on behind the scenes, like the workflows and security that we'll talk about. And so

2:36 for this video, I'll go through how Oracle AI Agent Studio works as a whole, but then also extrapolate lessons as we go through how everything is set up. And so I'll link to Oracle AI Agent Studio

2:46 in the description if you want to check it out. When we get started, it is just like working with our coding agent where we simply describe what we want to build. The important thing here though

2:54 is to again focus on the business outcome. What do we want the app to help us with in our role? And so for this example here, I say I want to build an app for manager coaching. I'm a director

3:05 level manager with a team of about 12 people and I want one view that is the agentic app of how engaged my team is with their development. And so a lot of different examples that we can build

3:15 here. There's other ones that you can check out as well, but this is a really good starting point just to take a look at how we can automate things with multiple agents in this application. And

3:23 so I sent in that prompt. I don't want to wait for it to build everything from scratch here. And so what I'm showing you now is the exact result I got from running that prompt earlier. And so

3:32 there's a lot of different components to the application here. And under the hood, everything that uses an agent, I'll show you some examples in a bit is running a workflow, right? So we have a

3:42 more deterministic process that is guiding how we use the agent to help us with the automations we have here. So it took a bit because the workflow was very comprehensive as you can see, a

3:51 wonderful answer, and we can ask any follow-up questions. So we still have the chat experience. We're not losing that in the Aentic app, but we also just have a lot more structure with all the

4:01 components that we have preconfigured here. Like, as another really quick example, if I scroll down to this bottom component, I can send a follow-up email to anyone who is not meeting their

4:09 current objectives. And so, again, calls a workflow under the hood to write the draft here. But the beauty of having this instead of a just chat application is that I have the component rendered

4:20 here, right in my dashboard where I can review the email, even edit it before I submit it. So I don't have to go over to another application. I have control all within the app. That's the entire idea.

4:30 It is one view to manage everything you need for a single business outcome. Like in this case, managing our direct reports. And take a look at the top right here because this is one of my

4:39 favorite parts of the application. So yes, there are agents running behind each one of these components, but there's also a different set of background agents that are running,

4:47 looking over everything we have here with direct reports and priorities and suggesting the next actions for me. So yeah, I can take a look at each one of these components if I want more detail.

4:57 But if I just want the high level, what should I do next? I can look here. So we have a suggestion from the agents to send an email to my direct reports. I can click in here to edit it before I

5:06 send and actually follow through on the action. So it helps us not just think about what we should be doing and our priorities, but helping us action on them as well. And now with that, the

5:15 most important thing here, if there's one takeaway you can get from this video, it's that whenever we invoke a large language model in an agentic app in any of these components here, it is

5:25 always going to be a part of a larger workflow to add more determinism. Like here's one example where we can see that we're injecting context at the start of this workflow run. We have code nodes

5:35 and conditionals. We're calling the LLM, right? Like it's very much an important part of the process, but it's not the only thing we're doing. like a lot of chat applications where you send the

5:45 input and it's given right to the LLM. For example, another thing we might need to do is change the format of the answer for how we want to display it in the agentic app, right? Just in general, we

5:54 have to have a lot more control than sending the request straight to the agent. So that's why we have the workflows that we run whenever we want to actually interact with an agent. And

6:03 I love this because if you're not aware, the entire industry is converging on this idea where you want to make your agentic workflows as deterministic as you possibly can because yes, LLMs are

6:13 very capable. They're flexible. They're able to look at a problem and figure out the steps to get there. But sometimes, especially on an enterprise level, you need to have certain guarantees like

6:22 context is always loaded in a certain way. You're calling the right LLM depending on the job with your conditional, your formatting things, right? like you have these hard

6:30 requirements where you can't give the agent the opportunity to get it wrong because it's just too missionritical or plain and simple you just want your workflows to be more reliable. This is

6:40 the most reliable way of doing things and even for AI coding workflows. I do this with Archon, my open source harness builder. In fact, my workflow builder there looks very similar to what we have

6:49 here in Oracle AI agent studio. So, I was able to jump into this right away and it just all made sense immediately because this is just the way to automate anything with LLMs now. And so in Oracle

6:59 AI agent studio essentially when you want to build an app you start by building the workflows. You figure out the specific automations that you want to bring into that single view and then

7:08 each component is going to leverage that workflow. Right? So clicking on different buttons here like this one I showed earlier directly invokes that workflow with what we have in the

7:16 dashboard as the input. And another thing that these workflows enable for us that's very important is bringing in tighter security controls. And so in Oracle AI Agent Studio, we have the idea

7:26 of a policy where we're able to create these certain functions that give agents very granular permissions and capabilities and then we're building these policies into our workflows. And

7:36 so let's say for this example, I'm building an agentic app for someone in the HR department. And as a part of that single view, they want an agent where they can ask about the leave policy. And

7:46 so what you can do in Oracle AI agent studio is you can just upload documents directly like in this case uploading a PDF document for the HR handbook. It takes that and it codifies all the

7:57 guidelines there into the functions that we're giving to the agent. Every single function has the source code generated test cases for validation output strict input. Right? The idea is overall we are

8:09 figuring out the exact capabilities the agent needs based on the input and then creating the functions for it. And so now as a policy that we've generated and tested here, we can use that in our

8:19 workflows. Now there's an important distinction I want to make here because a lot of people would try to solve this kind of thing with rag. So for the HR handbook, instead of going through this

8:28 workflow, turning it into code, a lot of people would just upload this to the agent and have it search pulling chunks from the document to answer questions on the fly. The problem is that is a

8:38 probabilistic answer. Now rag is good. it's going to get the answer right usually. But for something like this with benefits and vacation policies, usually is not good enough, right? We

8:49 need to take this document and codify it into tested code so the agent has structure and can get the answer right 100% of the time. And so for this workflow, we start with a classification

9:00 call to an LLM. So identifying is this a parental leave question or is it a vacation policy question. But before we call the LLM, we load the policy, right? We give the exact tools the agent needs

9:11 and nothing more and we scope it. And so within this workflow, we can also test the agent directly. So you'll see that we get JSON output here because this isn't the final form of the agent. This

9:22 is just the workflow that we're going to bring into some UI component in the Aentic app. We're just able to test it here as we are building out the workflow. And so we covered permissions

9:31 for the agent and the specific capabilities we're giving it, but also for the people using the Agentic app, you can manage the security of the individual workflows as well. So we can

9:40 scope who is able to run the workflow in the first place. Then also the identity that the workflow runs under as it goes through the automation. And there are of course a ton of other things that we can

9:49 tweak here like the LLM that we're using throughout the workflow, different triggers that we have for it because maybe we want things that are automated beyond just a user clicking something in

9:57 the Aentic app. It's just the app is the view for us, but then we can even expand that to, you know, automatically triggering things based on emails that come in, for example. The point is your

10:07 agentic app is your home base. It's where you go to get status updates from your agent to trigger workflows manually when you need to interact with your agent to ask it questions, everything

10:17 all in one place. And you can obviously expand the workflows and automations that you have here as well. Awesome. So, the last thing I want to cover quickly is how we can use our Aentic apps to

10:26 work with our agents in a more visual way, like we're looking at right here, cuz it's not really enough to just live in a chat platform like this 24/7. It certainly gets the job done a lot of the

10:37 time, but it's hard to always be in a place like this, right? We want our agent as much as possible to communicate things to us visually. So even producing artifacts like this slide decks, PDFs,

10:49 even if it's not something that we're going to use later, but just for our own faster understanding, this makes a big difference in our productivity. So especially when I want my agent to do

10:58 anything more comprehensive like a rundown, I don't want to just spew out a bunch of text. And so like we see what's built into Oracle AI Agent Studio right here. It outputs a PDF with the learning

11:09 risk summary going back to the first example of the Agentic app. And so I can go through the details here in a much cleaner way. Like you can imagine that if it gave me this, but just as a text

11:19 response, it'd be hard to go through that. And yeah, you can do it, but if you have to do that time and time again throughout the day as you're using your app, it just becomes a big mental drain.

11:28 And so I find that even just for my own sanity, I want my agents to be giving me output like this as much as possible. And hey, that's why I always build visualizations like this, these diagrams

11:38 to go along with my blabbing. And that's everything that goes into an agentic app and how you can use Oracle AI Agent Studio to get there. The main takeaway for this entire video is focus on the

11:49 business outcome. Right? Oracle's definition of an agentic app is building the app around the outcome instead of just trying to make the agent the entire thing. Now, keep in mind this is just an

11:59 intro video. I wanted to show you at a high level how we can build these agentic apps the way that Oracle implements them where business users can build everything here without being

12:07 technical at all. But I do also have some follow-up videos where I'll show how more professional developers can build these kinds of things directly in VS Code with coding agents. So

12:17 definitely stay tuned for those. And so if you appreciated this video and you're looking forward to more things on Agentic apps and Agentic engineering in general, I would really appreciate a

12:25 like and a subscribe. And with that, I will see you in the next