YT YouTube2Text

YouTube2Text → Transcripts → Agentic Apps: How Big Companies Are Actually Deploying AI Agents

Agentic Apps: How Big Companies Are Actually Deploying AI Agents — Transcript

by Cole Medin · 2,758 words · 391 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I focused a lot on building personal
0:02agents on my channel this year, like
0:03your second brain, coding assistants,
0:05even customer support agents. They're
0:07all the same where it's just a single
0:09agent you're building a use case around.
0:11You have that conversation with some
0:13kind of chat input where you give it
0:15requests and you read through the
0:16answers that it gives back. And these
0:19kinds of agents are fantastic. They run
0:20a lot of my business. But when I
0:22actually go into a company to integrate
0:24agentic AI, a lot of the times they need
0:26something more holistic. Now, especially
0:29because of the AI craze the last couple
0:30of years, it is very tempting for people
0:32building agents to make the agent the
0:35entire automation or application instead
0:38of it just being a tool within a larger,
0:40more deterministic workflow or
0:41application, which is actually what most
0:44people really need. They need what is
0:46called an agentic app. This is Oracle's
0:48official definition and I love it.
0:50Basically, with agentic apps, instead of
0:52thinking about the agent first, like
0:54what are we going to automate with this
0:55AI agent, we start by thinking about the
0:57business outcome, right? Like how are we
0:59going to help someone with part of their
1:01role, building the experience for that,
1:03and then once we have this mapped out,
1:05thinking about how we can build even
1:06multiple AI agents into the application
1:09to help with some of the automations.
1:11And honestly, this is just the most
1:12practical way to integrate agentic AI.
1:14If you focus on the business outcome
1:16first, that means you're much more
1:18likely to actually deliver value to the
1:20organization instead of just integrating
1:22agents for the sake of getting on the
1:24latest technology, which trust me, I see
1:26that way too often. And when your AI
1:28agents are simply a tool of a larger
1:30app, it gives you a lot more control,
1:32right? You're building more determinism
1:34because you're creating these workflows
1:36that use LLMs throughout the
1:37application, but it's not the agent
1:39driving the entire thing all the time.
1:41So, I hope that makes sense why we care
1:43about building these agentic apps
1:44focused on a business outcome. And so,
1:47in this video, I want to get very
1:48practical with you and show you what
1:49this looks like to go from a chat
1:52application like this, like this is my
1:53second brain, to an agentic app where we
1:56have many agents working under the hood
1:58to help us with a business outcome,
2:00right? Like this is helping us manage a
2:02role that we have in the business. Now,
2:04this is more enterprise focused, but
2:06even if you're building personal
2:07applications, there's a lot that you can
2:09learn from this. There's also a lot of
2:11different ways we can build these
2:12agentic apps. Now, out of all the
2:14platforms I've seen out there, Oracle AI
2:16Agent Studio does it the best. So, we're
2:18going to be showcasing Oracle AI Agent
2:20Studio throughout this, and I've also
2:22partnered with them to bring this video
2:23to you. But, just in general, there's so
2:25much we can learn from how Oracle AI
2:27Agent Studio builds these agentic apps
2:30and all the important things that go on
2:31behind the scenes like the workflows and
2:34security that we'll talk about. And so,
2:36for this video, I'll go through how
2:37Oracle AI Agent Studio works as a whole,
2:39but then also extrapolate lessons as we
2:41go through how everything is set up. And
2:43so, I'll link to Oracle AI Agent Studio
2:46in the description if you want to check
2:47it out. When we get started, it is just
2:49like working with our coding agent where
2:50we simply describe what we want to
2:52build. The important thing here, though,
2:54is to again focus on the business
2:56outcome. What do we want the app to help
2:58us with in our role? And so, for this
3:01example here, I say I want to build an
3:02app for manager coaching. I'm a
3:04director-level manager with a team of
3:06about 12 people, and I want one view
3:08that is the agentic app of how engaged
3:10my team is with their development. And
3:13so, a lot of different examples that we
3:14can build here. There's other ones that
3:16you can check out as well, but this is a
3:17really good starting point just to take
3:19a look at how we can automate things
3:20with multiple agents in this
3:22application. And so, I sent in that
3:24prompt. I don't want to wait for it to
3:25build everything from scratch here. And
3:27so, what I'm showing you now is the
3:28exact result I got from running that
3:30prompt earlier. And so, there's a lot of
3:32different components to the application
3:34here, and under the hood, everything
3:36that uses an agent, I'll show you some
3:38examples in a bit, is running a
3:40workflow, right? So, we have a more
3:42deterministic process that is guiding
3:44how we use the agent to help us with the
3:46automations we have here. So, it took a
3:48bit because the workflow was very
3:50comprehensive, as you can see. A
3:51wonderful answer, and we can ask any
3:53follow-up questions. So, we still have
3:55the chat experience. We're not losing
3:57that in the Agentyc app, but we also
3:59just have a lot more structure with all
4:01the components that we have
4:02preconfigured here. Like as another
4:03really quick example, if I scroll down
4:05to this bottom component, I can send a
4:07follow-up email to anyone who is not
4:08meeting their current objectives. And so
4:10again, calls a workflow under the hood
4:12to write the draft here, but the beauty
4:15of having this instead of a just chat
4:17application is that I have the component
4:20rendered here right in my dashboard,
4:21where I can review the email, even edit
4:23it before I submit it. So, I don't have
4:25to go over to another application. I
4:27have control all within the app. That's
4:29the entire idea. It is one view to
4:31manage everything you need for a single
4:33business outcome. Like in this case,
4:34managing our direct reports. And take a
4:37look at the top right here, because this
4:38is one of my favorite parts of the
4:40application. So, yes, there are agents
4:42running behind each one of these
4:43components, but there's also a different
4:45set of background agents that are
4:47running, looking over everything we have
4:48here with direct reports and priorities,
4:51and suggesting the next actions for me.
4:53So, yeah, I can take a look at each one
4:55of these components if I want more
4:56detail, but if I just want the
4:57high-level, what should I do next, I can
4:59look here. So, we have a suggestion from
5:01the agents to send an email to my direct
5:03reports. I can click in here to edit it
5:05before I send, and actually follow
5:08through on the action. So, it helps us
5:09not just think about what we should be
5:11doing in our priorities, but helping us
5:13action on them as well. And now, with
5:14that, the most important thing here, if
5:16there's one takeaway you can get from
5:18this video, it's that whenever we invoke
5:20a large language model in an Agentyc
5:22app, in any of these components here, it
5:24is always going to be a part of a larger
5:27workflow to add more determinism. Like
5:29here's one example where you can see
5:31that we're injecting context at the
5:32start of this workflow run. We have code
5:35nodes and conditionals. We're calling
5:37the LLM, right? Like it's very much an
5:39important part of the process, but it's
5:40not the only thing we're doing, like a
5:43lot of chat applications where you send
5:44the input and it's given right to the
5:46LLM. For example, another thing we might
5:48need to do is change the format of the
5:50answer for how we want to display it in
5:52the Agentic app, right? Just in general,
5:54we have to have a lot more control than
5:55sending the request straight to the
5:57agent. So, that's why we have the
5:59workflows that we run whenever we want
6:01to actually interact with an agent. And
6:03I love this because if you're not aware,
6:05the entire industry is converging on
6:06this idea where you want to make your
6:08Agentic workflows as deterministic as
6:10you possibly can. Because, yes, LLMs are
6:13very capable. They're flexible. They're
6:15able to look at a problem and figure out
6:16the steps to get there. But, sometimes,
6:19especially at an enterprise level, you
6:20need to have certain guarantees. Like,
6:22context is always loaded in a certain
6:24way. You're calling the right LLM
6:25depending on the job with your
6:27conditional, your formatting things,
6:29right? Like, you have these hard
6:30requirements where you can't give the
6:32agent the opportunity to get it wrong
6:34because it's just too mission-critical.
6:36Or, plain and simple, you just want your
6:38workflows to be more reliable. This is
6:40the most reliable way of doing things.
6:41And even for AI coding workflows, I do
6:43this with Arkon, open-source harness
6:45builder. In fact, my workflow builder
6:47there looks very similar to what we have
6:49here in Oracle AI Agent Studio. So, I
6:51was able to jump into this right away,
6:53and it just all made sense immediately
6:54because this is just the way to automate
6:56anything with LLMs now. And so, in
6:58Oracle AI Agent Studio, essentially,
7:00when you want to build an app, you start
7:02by building the workflows. You figure
7:03out the specific automations that you
7:05want to bring into that single view, and
7:07then each component is going to leverage
7:09that workflow, right? So, clicking on
7:11different buttons here, like this one I
7:12showed earlier, directly invokes that
7:14workflow with what we have in the
7:16dashboard as the input. And another
7:18thing that these workflows enable for us
7:20that's very important is bringing in
7:22tighter security controls. And so, in
7:24Oracle AI Agent Studio, we have the idea
7:26of a policy. We're able to create these
7:28certain functions that give agents very
7:30granular permissions and capabilities,
7:33and then we're building these policies
7:34into our workflows. And so, let's say
7:36for this example, I'm building an
7:38Agentic app for someone in the HR
7:40department. And as a part of that single
7:42view, they wanted agent where they can
7:44ask about the leave policy. And so what
7:46you can do in Oracle AI Agent Studio is
7:48you can just upload a documents
7:50directly. Like in this case, uploading a
7:52PDF document for the HR handbook. It
7:55takes that and it codifies all the
7:57guidelines there into the functions that
7:59we're giving to the agent. Every single
8:01function has the source code generated,
8:03test cases for validation, output,
8:06strict input, right? The idea is overall
8:08we are figuring out the exact
8:10capabilities the agent needs based on
8:12the input and then creating the
8:13functions for it. And so now as a policy
8:16that we've generated and tested here, we
8:18can use that in our workflows. Now,
8:20there's an important distinction I want
8:22to make here because a lot of people
8:23would try to solve this kind of thing
8:24with rag. So for the HR handbook,
8:27instead of going through this workflow,
8:28turning it into code, a lot of people
8:30would just upload this to the agent and
8:32have it search pulling chunks from the
8:33document to answer questions on the fly.
8:36The problem is that is a probabilistic
8:39answer. Now, rag is good. It's going to
8:41get the answer right usually, but for
8:43something like this with benefits and
8:45vacation policies, usually is not good
8:47enough. Right? We need to take this
8:50document and codify it into tested code
8:52so the agent has structure and can get
8:54the answer right 100% of the time. And
8:57so for this workflow, we start with a
8:59classification call to an LLM. So
9:01identifying is this a parental leave
9:03question or is it a vacation policy
9:05question? But before we call the LLM, we
9:08load the policy, right? We give the
9:10exact tools the agent needs and nothing
9:12more and we scope it. And so within this
9:14workflow, we can also test the agent
9:16directly. So you'll see that we get JSON
9:18output here because this isn't the final
9:20form of the agent, this is just the
9:22workflow that we're going to bring into
9:24some UI component in the agentic app.
9:26We're just able to test it here as we
9:28are building out the workflow. And so we
9:30covered permissions for the agent and
9:32the specific capabilities we're giving
9:34it, but also for For people using the
9:36agentic app, you can manage the security
9:37of the individual workflows as well. So,
9:40we can scope who is able to run the
9:42workflow in the first place, then also
9:43the identity that the workflow runs
9:45under as it goes through the automation.
9:47And there are of course a ton of other
9:49things that we can tweak here like the
9:50LLM that we're using throughout the
9:52workflow, different triggers that we
9:53have for it because maybe we want things
9:55that are automated beyond just a user
9:56clicking something in the Agentyc app.
9:58It's just the app is the view for us,
10:00but then we can even expand that to, you
10:03know, automatically triggering things
10:04based on emails that come in for
10:06example. The point is your Agentyc app
10:08is your home base. It's where you go to
10:10get status updates from your agent, to
10:12trigger workflows manually when you need
10:14to, interact with your agent to ask it
10:16questions, everything all in one place
10:18and you can obviously expand the
10:19workflows and automations that you have
10:21here as well. Awesome. So, the last
10:23thing I want to cover quickly is how we
10:25can use our Agentyc apps to work with
10:27our agents in a more visual way like
10:29we're looking at right here. Cuz it's
10:31not really enough to just live in a chat
10:33platform like this 24/7. It certainly
10:35gets the job done a lot of the time, but
10:37it's hard to always be in a place like
10:40this, right? We want our agent as much
10:42as possible to communicate things to us
10:44visually. So, even producing artifacts
10:47like this, slide decks, PDFs, even if
10:49it's not something that we're going to
10:50use later, but just for our own faster
10:53understanding, this makes a big
10:55difference in our productivity. So,
10:56especially when I want my agent to do
10:58anything more comprehensive like a
10:59rundown, I don't want to just spew out a
11:02bunch of text. And so, like we see
11:04what's built into Oracle AI Agent Studio
11:06right here, it outputs a PDF with the
11:08learning risk summary. Going back to the
11:10first example of the Agentyc app. And
11:12so, I can go through the details here in
11:14a much cleaner way. Like you can imagine
11:16that if it gave me this, but just as a
11:18text response, it'd be hard to go
11:20through that. And yeah, you can do it,
11:22but if you have to do that time and time
11:24again throughout the day as you're using
11:25your app, it just becomes a big mental
11:27drain. And so, I find that even just for
11:30my own sanity, I want my agents to be
11:32giving me output like this as much as
11:34possible.
11:35And hey, that's why I always build
11:36visualizations like this, these diagrams
11:38to go along with my blabbing. And that's
11:40everything that goes into an agentic app
11:42and how you can use Oracle AI Agent
11:44Studio to get there. The main takeaway
11:46for this entire video is focus on the
11:48business outcome, right? Oracle's
11:50definition of an agentic app is building
11:53the app around the outcome instead of
11:55just trying to make the agent the entire
11:57thing. Now, keep in mind, this is just
11:59an intro video. I wanted to show you at
12:00a high level how we can build these
12:02agentic apps the way that Oracle
12:04implements them. Where business users
12:05can build everything here without being
12:07technical at all. But, I do also have
12:10some follow-up videos where I'll show
12:11how more professional developers can
12:13build these kinds of things directly in
12:15VS Code with coding agents. So,
12:16definitely stay tuned for those. And so,
12:19if you appreciated this video, you're
12:20looking forward to more things on
12:21agentic apps and agentic engineering in
12:23general, I would really appreciate a
12:25like and a subscribe. And with that, I
12:27will see you in the next video.

About this transcript

This page contains the full transcript of Agentic Apps: How Big Companies Are Actually Deploying AI Agents by Cole Medin, generated from the public captions YouTube serves with the video. The transcript has 2,758 words across 391 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.