DON'T Build AI Automations, build agentic workflows! (Claude Managed Agents)

Guys, I'm super excited to get into this video today because we're gonna go over several very important foundational concepts when it comes to building AI workflows that I don't think anyone else in the space is talking about right now. Before we get into that, I wanna talk about Anthropic's new managed agents platform. So this is a, this is an option inside of Claude's console where you can actually build out full agentic workflows directly using Claude. Using Anthropic. It is a very interesting idea. It's kind of like their version of N8N or Zapier. And I can see this as someone who is continuously like building automations. I think that this is gonna become something every single business owner needs to understand in 2026 because what's happening over here, and if we look at some of the quick start examples that they share, you can see there's agents that are gonna start helping with checking Notion, checking Slack, doing feedback mining, Incident Commander, Field Monitor. These are all skills that people are doing that we actually are paying people to do for us. So the second that the agents can start to build this out, it is going to be a new world. And that's why as a business owner, you need to be on top of the stuff and paying attention to this stuff. So before we get into building with tools like managed agents, we need to talk about the foundations of this. I wanna show you and really get you thinking about the meta skills, how to, how to understand first principles of working with AI so that you can get the best outcomes. And now why would you wanna listen to me on this? Well, my name's Jason Edward. I run a 7-figure marketing agency and we've been able to automate more than 50% of our operations by using these agents, by using workflows and systems. And I wanna show you how you can do it for yourself as well. So we're gonna get into a couple things in this video. Uh, first I'm gonna go over an analogy that was officially, uh, dropped by Anthropic. Uh, we're gonna talk about two different types of AI that you need to understand in order to build workflows. We're gonna go into what are agentic workflows and why I feel this is the most important concept for you to master in 2026. Uh, how to build in pipelines. And finally, we're gonna get into some real world applications, some real world examples so you can start to see how easy it is to apply workflows and agents into your actual business so that you can start to benefit from saving time. So with that said, let's talk about what's going on here with managed agents in Claude Code. So Anthropic released a blog post and, uh, they were talking about how they went from, they went into a new paradigm of how they were thinking about these agents. So previously, you know, you had what was called a harness. So Claude Code, Claude Cowork, these are platforms where you can add tools, you can add MCP servers and different resources that the agent has access to. And then you can also create skills for the agent and it can create code, it can write code, it can run commands. But the problem is that as you start to stack more and more capabilities onto it, you would naturally run into issues with your context window. It would start to break. Maybe the system isn't working as it's supposed to. And it's very, very difficult to debug it because you have one harness that has so many different responsibilities across your business. The problem is also if this one goes down, everything goes down and that led to a couple problems. So what they've done with managed agents is they've separated the harness from the actual execution, from the actual workflows that they're building. And as they actually mentioned in their blog post, they've decoupled the brain from the hands. So they started by saying, we originally coupled everything into one container and we ran into the old infrastructure problem. We've adopted a pet. So there is this analogy in, in software development called pets versus cattle. And you can see this picture over here. Basically, you can either think about building a pet, is a hand-named, hand-tended individual where you can't afford to lose it, and cattle, which are interchangeable. And in our case, the server became the pet. If a container failed, the session was lost. If a container was unresponsive, we had to nurse it back to health. The solution we arrived at was to decouple what we thought of as the brain from, like Claude and his harness, from the hands, which is the sandboxes and tools that perform the actions. Each became an interface that could make a few assumptions about others and each could fail or be replaced independently. So with that said, the key, the key here is interoperability. This is the keyword. And if you don't know what that means, interoperability is a critical ability of systems, devices, organizations to exchange and use their data seamlessly. So that's what we're going to be diving into today. We're going to talk about how you can build systems in your business that are reliable, where even if one part of it stops working, you don't, uh, the, the whole system doesn't fail. And then also it's very easy to debug where the problems are. So next, let's understand what are the two main types of AI. Keep in mind, I'm gonna keep this extremely simple, uh, in this video. If you're in the comments saying like, look, this is not entirely accurate, like point being, there's really only two that you need to think about here. We've got our language models, so ChatGPT, Gemini, Claude, Groq. So these are the ones that you're using to create text. So these are LLMs. They are inferring meaning from language, and then they're also able to reason through the language. From there, uh, another thing that they can do is math. So they're very good at, you know, linear equations and understanding formulas and than writing code. So any application, uh, this is an important concept. If you, if you see writing code and you think, well, I'm not a coder, I don't know how to do code. All you need to understand is that code is what runs everything in our digital world. So YouTube, uh, this video that I'm, that I'm, that I'm recording for you right now, it's all being run on software. And the key concept to understand is if, if the AI, if the language model can write code, it can also understand code. And if it can understand code, it means it can manipulate and mold the different services and softwares that you're using on a daily basis. So that's a huge paradigm shift. And once you understand that, it becomes very powerful. So when I, uh, put here like APIs, computer use, browsing, these are all key, uh, abilities or attributes of language models, which make them extremely powerful to work with. Then they also have vision. So another really powerful aspect is you can give it an image and it can tell you everything about the image and it can reason, it can use all of these things simultaneously with the image provided. So that's the key concept to understand. We've got language models, then we've got our image and video models. So this is like Nanobanana, V03, Sora, SeedDance. So what these can do is they can generate images. You can either do text to image, meaning you just say what you want and it creates a picture for you, or you can do image to image, in which case you give it an initial image and it can generate a new one or edit the image how you want it to, to be. Then we have generating video. So this one is what, text to video, you can speak and it generates a video, image to video, and video to video. So really simple. And why I mentioned that you need to understand these two types is because when you're building your workflows, the various different automations, you can start to use these tools together. You have access to generate images. So what does that mean? Marketing content, your social media posts, maybe a certain, uh, product photography that you want in, in your business, uh, generating videos. So this could be explainer videos. The, these tools are continuously getting better. And so I wanted you to just have the foundational understanding because as they develop in the next 6 months, in the next year, you can understand and you can immediately be on the cutting edge when the new versions come out. Um, on the language model side, it's like if you can understand that they can generate text, they can do math, and they can write code, you can start to see how that fits together into different workflows you can build. Cool. Let's move on to agentic workflows. So like I mentioned over here, uh, Agentic is where you, like the way you wanna think about this, right? So first of all, what is an agent? An agent is something that can run and basically just do work on your behalf. Like the idea is long horizon task. So it can handle undefined problems. It adapts to new information. It can decide what steps to take based on inputs that you provide. So if you look at a tool like Claude Code or Claude Cowork, This is a great example of agents because it can take your input, your prompt, or any, you know, link that you send it. It can go and do web research. It can then reason through that. It can write some code for itself. It can ask you if you, if it needs anything or if it has any questions and call tools. So you provide it with different tools such as like access to your email, access to your Slack, and it can go and execute tasks on your behalf. Now, this is great. The problem comes in where the agentic, uh, the agentic code, as we mentioned, like the running with an agent, it, it's not as reliable. It is kind of random. Sometimes it might do things differently depending on how you give it the inputs. And this is not great for building workflows. What we want in workflows is we want to have predictable deterministic systems, something where, you know, if you provide this input, you're going to get this output. This is also very important when it comes to debugging, because if you just rely on an agent, if you're relying on an agent that has just access to various different tools, it might not do things the way that you want it to do things every single time. So we need to understand how workflows play a role here. And you can see we have this diagram. This is a classic workflows diagram: input, process, output, and storage. So let's actually take a look at what, what this workflows concept, uh, how you can actually apply this to, to agentic workflows. So the most important thing to understand here is that we are always gonna go through the same kind of core phases. So we've got your trigger. So that is how does this entire thing kick off in computer programming, in, in working with agents, we need to know what is gonna start the agent because Of course it's spending, it's spending tokens and credits, so it's not gonna be running 24/7. It needs some sort of trigger to get it started. Now that could be a manual prompt, such as if you're just speaking to Claude Cowork, you can tell it like, hey, I have this task for you and then it will begin working on it. Or a scheduled reminder. So hey, every day at 8:00 AM I want you to go and do web research. That's an example of a reminder. And then finally is external events. So this would be like a webhook or when some other event happens, someone submits a form, a calendar gets created, an email gets sent. You can get your agent to like wake up and start working with the workflow you have in mind. Next is ingestion. So this is, uh, you could, this has many different names. You could also call this your inputs, but the idea that I want you to understand here is for the agent to execute tasks on your behalf, it first needs to get the right context. So what we're doing is we need to have a good way to give it access to everything that it needs to make the best decisions. So if we give it the right tools, then it can reason through those tools and actually think about it. So where do we get the data from? This is gonna be your meetings, your emails, Slack messages, web browsing. So setting it to go to specific websites or to just scan scan the internet for information that it needs. Company documents, so your SOPs, your, your various different files that you, that you're creating. Analytics, so this could be your Google Analytics, your Stripe data, your Shopify data, any, any analytics platform that you're using in your business to keep track of your, your metrics. Spreadsheet data, so your financials, maybe your accounting, your, your P&L. This is all stuff that you can give your agent to reason about. API. So this is really where it just expands to everything because every single app should in 2026 have an API. And that means that you can kind of get your agents to use your computer and use your software just like you could. Very, very powerful. And finally is like internal memory. So this is something which is also becoming more well-known concept. It's like we need the agents to be able to remember, remember information that we told it before And that again is just an input that we're using to control its context. So this is really the, the art of context engineering. And what I will say is all of these concepts like triggers, ingestions, transformation, actions, these all require videos, uh, just on themselves. Like these are extremely, uh, important concepts. And I want to go into more depth on these in future videos, but at least now we can get the overall picture of this, right? Next we have transformation. So once we've ingested the inputs that we need, next your agent needs to actually reason and transform this data. So this is where I put, uh, the word intelligence because while the ingestion part of this does not necessarily need an AI, it is going to be built by your AI. But once it's set up, it is really just code that's running. On a specific trigger. The real intelligence, the, the actual reasoning steps come from, uh, this part over here. So once we've got the data in a format that the agent can understand, uh, now we're getting it to think logically through the relationships. You know, if you give it meeting data and you give it your spreadsheet data, how can it infer relationships between these two data points and actually come up with the right decisions to make based on that? This is where, uh, a lot of the work of agentic engineering and also building these agentic workflows needs to be done because you're trying to get this agent to think through how to perform a task and how to reason the same way that your team would do it, or the same way that you would do it with the, the same information. So again, we provide it with consistent inputs and we wanna make sure that it thinks through these inputs in the right way. To get you the output that you're looking for. So what are the different stages of transformation? Well, it could be reasoning, enrichment. So that means understanding the relationships and then, and then basically just figuring out like what data is the most important extraction. So finding and surfacing critical information within the data. So being able to, you know, just pull out someone's name, their age, their location, et cetera, from the data and then fill that into a spreadsheet. And finally is like conversion. So again, these are, these are all like meta points. And, uh, you know, we, we will build on these in future videos, but at least if you understand that this is what you can use your intelligence for, um, yeah, conversion is, is being able to change it from like text to image or image to text or markdown to HTML. Like these are things that can be done in the transformation stage. Finally, we've got your actions. So once you've transformed your data, what's gonna happen is you probably want the agent to send to, to, to do something with that information. So that could either be creating a document, sending you an email, letting you know that it's been created, adding an event to your calendar, uh, creating and editing tasks in your project management tool, creating new spreadsheets, building slide decks. Updating your CRM, updating the memory system, as we spoke about, and also maybe even verifying with you. So asking for your inputs before that it continues. So again, these are actions and a key concept with what I wanna show you over here. Uh, I put these kind of like refresh arrows. The reason why is because this process of trigger ingestion transformation action is not really a linear process. Especially as we start to get more into agentic behavior, it is kind of like a back and forth process and your actions can also become your inputs. So if you take some sort of information like, uh, web, web data and documents that, that, uh, that your company has, you can transform that information. And the action could be creating it into an email or creating a spreadsheet. And that spreadsheet could then trigger some additional transformation that happens. So it's like, okay, you first generate 50 columns in a spread— or 50 rows in a spreadsheet. Now, how you can use that again to reason through that, to further come up with new, new insights and then use that to apply in your business. So hopefully you can understand over here, this approach of ingestion transformation actions is really, it's, it's kind of like more of a circle than a linear approach. I think that's just the key paradigm I want you to understand. Awesome. So with that said, let's talk about the very next concept over here, which is once we've, once we understand the meta of what we're really building with these agentic workflows, we need to talk about pipelines. So building in pipelines. So when I say pipelines, what I'm getting at is step-by-step, a step-by-step process, a repeatable form formula that you want to, uh, have your agents execute. And this is really the heart of, uh, how you can get consistent outputs from these, from these tools. I have many friends that have already told me now they're using Open Call or using Cloud Cowork, and they are frustrated because they set up some sort of skill. They set up a, a process they wanted the agent to do, and then it worked perfectly for one day. And then a few days later they come back to it and suddenly it's not working anymore, or it's broken, or the agent did the wrong thing and they're just frustrated and it's difficult to debug it. So if we start to think in pipelines, that kind of frustration is not gonna happen anymore. I'll tell you why. First thing we wanna understand here is whatever output, whatever outcome you're looking for, you know, Alex and Mozi gave the example that every single role in your company You need to understand what are their core responsibilities and then start to think, how can you automate or enhance or like make more efficient each one of those responsibilities? So you want to define a specific outcome, um, that maybe your team member is doing for you. Uh, this could be things such as like creating documents or creating spreadsheets or being able to go look up information. We're going to get into the real examples at the, at the end of this video. Uh, but most important thing is whatever outcome you have in mind for your automations, you can then reverse engineer that into the steps, the repeatable steps that need to happen, starting with the inputs. So, you know, what key information do you need to provide for the agent to go off and start to actually execute these tasks for you? Right. So with that said, the next thing is you need to understand the anatomy of each of these steps. In order for you to build complex workflows that are, that have multiple chains of actions that you want the agents to take, it's very important that you, that you realize, um, each, each step is really just a process of inputs and outputs and processing. So you pro— you give it your inputs, it's gonna run some function or some, some formula or some reasoning step, and then it's gonna give you an output. And the key mindset here when it comes to building pipelines is you need to have a validation, uh, an evaluation function. So you wanna be able to know what is the outcome of that step that you're looking for, and then being able to test this within just the step itself. So the idea is instead of giving the agent a very complex task in one shot, and then just hoping, like praying that it's gonna get it right from end to end, you need to break the task down into much smaller subtasks. And then you can actually just start to test each of those individual steps. And each step, uh, working correctly means the whole overall kind of workflow will work correctly. So the idea is, is, is really just break it into steps, create the right inputs for each step, check the outputs, and then work with the agent and continue to iterate on each step until you're happy with the end result. Right. That is really the key foundational principle here. And from that, once you understand this anatomy, you can see how each step could be broken down into like many different sub-steps. So if you have, I want to give you a good example over here, but like, let's say you wanted the agent to go do research for you. In order to do research, you need to understand, okay, first of all, what does it need to do research about? So you need to get it to ask you questions. Then where is it going to go to do the research? You need to figure out what are your sources? Is there specific channels? Is there specific websites? That's the, the, that's a step in the process that the agent needs to understand. Then it becomes, okay, how is it going to go to those places? Do we have the right scraping tools or the right API key set up for it to be able to do that? And then once you have that, it's like, okay, well, let's say you get the information. How are you going to organize the information? And eventually you get to the point of outcome that you want, which is that you get back a research document. So that's a, that's a way that you can start to think about this. And again, we just want to break everything down into the smallest possible pieces. This is really the heart of many concepts in computer engineering is it's just a lot of individual functions that are built together to create a higher level of order, um, within the, within the machine. Right. And yeah, this just goes on and on. So you can just keep breaking it down into step 1, step 2, step 3. Et cetera. Awesome. So now that we've talked about pipelines, we can get to the final examples here. And these are the real-world examples of how you can actually apply this to your life, how you can apply these, uh, concepts of agentic workflows to your business. So the first one is just automated industry news that can be sent to you every morning. If you haven't seen, I actually already made a video, uh, which I'll link to in the top of this. Where we talk about how to build a super simple automation where every day your agent can send you an email at 8:00 AM where it goes through YouTube, blog articles, news sources, social media, X, Reddit, Instagram, and it pretty much compiles all the information related to any topic you want. And it extracts that information and reasons through it and then sends you an email. So you can see over here for each one of these examples, I, I use the exact same framework we spoke about of workflows. So you have your trigger, when it, when it starts, the ingestion, so where it gets its input from, where it gets the data from, how it transforms that, where it just collates, collects it all, and then, and then, uh, creates a, a document with it. And finally, it sends the email. The next one is, uh, financial analysis. So here's another example. We can say at 8:00 AM it triggers the agent. What it can do is it uses the API, it goes and fetches the latest financials from Stripe. It fetches your inventory data from a different, from a spreadsheet where you're managing that. It queries your bank and it gets the API to get transactions and balance. It's, the idea is it's doing the same thing that you would do in your business and you just need to provide it with the right inputs again. And from there it can extract the information that it needs. It can reason through that and it will from there be able to take the actions that you have trained it to do. which in this case could be creating some form of document and then sending an email to yourself or to a CFO, and then also updating its memories. So this could be a great example of some form of financial analysis agent, uh, intelligent agent where you could ask it about how your finances are doing. And because it has the way to get the information and, and use that, it can start to, uh, you know, do this for you on a recurring basis. Great. Uh, next one. So update your company's SOPs. So this is the one that I've actually been using myself. Uh, super powerful. If you have a team member or you're doing training with your team on a regular basis, you might be doing calls and you could take that call recording and you could literally just have a prompt where you say to the agent like, hey, take this call and build out some core SOPs. So what it will do in the ingestion step, it fetches the current training documentation. So maybe it looks at your existing SOPs that you've already created, and then it takes the call recording. It'll extract the information again, go through this transformation and then edit the existing SOPs that you have with new information and maybe summarize it and send you an update of, you know, what it's actually done inside of your SOPs. So again, that's like a repeatable process that we've built, uh, sales prospect enrichment. So this one is a slightly different trigger. Uh, let's say you have a form submission. So every time someone submits a form, it will go and it'll read through the URL that they submitted. It'll look at their Meta ad library. So you can have an automation set up that it literally just goes and finds their current ads. And then it also goes and finds their LinkedIn. This is again, just examples. From there, again, you are transforming that data. So it takes all those inputs. You can then create a preparation template. So if you're gonna do a, a sales sales call, you could prepare for the initial call with the different information. So instead of you having to go and do these steps yourself, you just build an agent that can do that for you. And then maybe sends that to your sales team on Slack. It could also create email templates. So follow-up email templates that are personalized to whoever submitted the form. Again, just, I want to get your mind thinking about how you can start to use this in your business. And number 5, finally, is product market research. So, Here's a really cool idea. And the idea here is with loops. So let's say you had a spreadsheet with 50 product ideas, right? For each one of these, you could build a workflow where it takes the idea. It then goes and researches suppliers from various different websites. And it finds quotes for the manufacturing of the product and also shipping. You can also at the same time, find competitors, others that are selling the same product, make a list of the costs, and the, you know, the specific meta information that you need. And then again, reason through that information. So basically extract what it thinks is important from that and decide what, uh, you know, what to do with that information and then potentially update the same spreadsheet. So by doing that, you're able to actually have a process where it's, it's doing the step of a human. It's taking very, very raw inputs and it's thinking about how to apply them strategically so that you don't have to spend your time doing that. So again, those are 5 real-world examples. I'm so stoked to make more videos about all of these concepts. And this is really just the foundation of what I'm getting into. So if you like this video, I really appreciate your support. If you would subscribe for, uh, stay tuned for my next videos, I am getting better at making videos in general. Um, so this, this is only going to get better. And I really hope to, uh, yeah, just share more with you and to learn together as we get into this crazy world which we're in right now of agents and workflows. It's crazy time. It's such a crazy time to be alive, but I hope that you enjoyed this video. Thanks so much for watching. And I look forward to seeing you in the next one. Bye.