Agentic Workflows Just Changed AI Automation Forever! (Claude Code)

AI automation is about to change forever. Most workflows today only do exactly what you tell them, nothing more. But Agentic workflows are different. You basically give them an outcome, not just a flow, and they figure out the steps by themselves. So, in this video, I'm going to explain the shift, and then I'm going to build one live so you can see just how practical it already is. So, let's first take a look at the difference between the new Aentic tech versus the old way of doing things. So, what does Agentic actually mean? Well, if you've been building an Enten for a while, you know the drill. You drag a node, you configure it, you connect it to the next one, you test the flow, you hit an error, you read the error message, you figure out what went wrong, you fix it, and then you test it again. Which means every API connection, every variable, every condition, that's all on you. But with Aentic, it's a bit different. Instead of telling the system how to do something, you just tell it what you want in plain English. So you explain where the data is coming from, what needs to happen, and then where it should end up. The agent then figures out the rest. So just think about it like you're hiring an actual human developer. You wouldn't go in there and explain every single line of code. You would just explain the outcome that you're looking for, what problem you're trying to solve, what tools you're working with, and what the end result should look like. And then they'd go build it for you. So, here's the thing. If you can't explain clearly what you want, then how could you expect a human or an AI agent to actually build that? So, there has to be a clear scope. You don't need to know how to code, but you do need to know how to communicate your plan clearly. And the good news is you can actually use the agent to just brainstorm. So, if you tell it roughly what you're trying to do, and it will ask you all the questions that it needs answered in order to actually build a robust solution. And that's the huge shift because you can go from being the person who constructs every piece of the puzzle to being the person who just defines what do you want the puzzle to look like. So there are four main changes that I need you to understand in order to build an agentic workflow with me. So the first one is self-healing. This is probably the biggest and one of the coolest ones because if you build something in Nitn when the workflow breaks you read the error message you tweak the node you test it again and if it breaks again you repeat. So that whole debugging loop eats up a lot of time. Now if you're building an agentic workflow the agent handles that loop for you. Imagine every time your workflow broke, a smart assistant was sitting right there next to you just reading the error, testing a fix, and only calling you in if it was really, really stuck. So, the agent tries something, runs it, checks what happens. And if something broke, it figures out why, edits its own code, edits its own instructions, and updates everything so that it doesn't make that mistake again. And here's what's really wild. You don't have to read or write any of that code yourself. You just explain what good looks like, and when you need to approve the changes that it suggests. All right, number two is natural language control, but for real this time. Because you've probably seen tons of tools that let you build with natural language. Lovable, Bolt, Lindy, even Nadn has an AI workflow builder. You basically type in a prompt, it spits out a workflow or a website. But if you've actually tried using these for productionready systems, you know that usually they require a lot of manual cleanup or a lot of back and forth. They'll get you maybe 60 or 70% of the way there. But this new generation, it feels a lot different. Instead of just taking your prompts and building, it first interviews you to make sure that it knows everything it needs. So, it'll ask you clarification questions like we talked about before it tries to write a single line of code. It'll ask you questions like who are the users? How many times should this run? What tools are you using? What happens if X happens? What should happen if Y happens? A lot of times when I'm working with these agents, they're asking me questions that I wouldn't have even thought of until a month later when we run into a problem. And then once it has all that information, it builds the systems and it wires it right into your own text stack already. So once that's built, natural language becomes the remote control that you can use for the whole automation. So you can have V1 ready and you can say h actually can you make that a little bit faster or can you make it cheaper or can you add a manual review step right here or can you log all of those outputs in this Google sheet and it will just go ahead and do it. Now of course you can't just completely like vibe code these automations. You still have to do a little bit of thinking but it will ask you the right questions and guide you. Now something else I think is really cool is that you can have multiple agents running simultaneously. So if I asked 10 of the best AI engineers that I knew to build me a solution, I'd probably get 10 different versions even though all 10 of them work. And sometimes the only way to know which one's best is just by testing them all out. So you could be brainstorming a process with an agent and you could say, "Please suggest the five best approaches for this, you know, workflow." And then instead of having to like guess which one you think you should try, just say, "Okay, cool. Agent one, try a method number one. Agent two, try two, three, try three," and so on. And then you can step away, grab a cup of coffee, come back when they're all done, and then stress test all of those solutions to see which one's like the cheapest, the fastest, and the highest quality. Number three is about security. So, when you think about the fact that you're building code that you don't really know what it's doing, you have to think about security. And these agents are writing and editing real code. And I know that a lot of you don't want to know how to read that or don't know how to read that. And I don't either. Just don't worry, you don't have to. Because these same large language models that are generating that code can also constantly review it for security problems and vulnerabilities. And they can do that on every single change, not just once. So, think about it like you're having a security obsessed developer sit right next to you while you're building your automation. And every time something changes, they double check. Are the API keys hidden? is sensitive data being logged somewhere that it shouldn't be. This can happen automatically as part of that self-healing loop. So, every time the agent proposes an edit, it reviews its own work, it makes the changes, and then it makes sure that everything's still secure. And of course, you can build in all of these guard rails with natural language. So, you can just say something like, "Never send customer phone numbers to any third party tool or always make sure to stop this workflow if you ever exceed $5 of usage with this API." Things like that. Your job is just to say what must never happen. And then the system's job is to actually enforce that. All right. Fourth is instant API and MCP integrations. So if you've built workflows in Naden, you know the pain of API authentication, headers, JSON bodies, query parameters. You spend half your time just reading API documentation and then getting the shape of the request to be right. With Aentic Workflows, that pain pretty much disappears because you just say what tool you want to connect to. So, you know, get my Fireflies transcripts and then push that into ClickUp and then send me a Gmail summary. All you have to do then is go in and grab your API keys and then the agent figures out the rest because instead of you having to read those, you know, 10 pages of API documentation, the agent can just search the web or even go to the URL you give it and research it on its own. Now, MCP is basically like an app store for these agents. They're tools that are pre-wrapped up and documented so that agents know what's available and how to use each of them correctly. And if a tool doesn't have an MCP yet, then the agent can just once again go to the API documentation and then make the right request there. And because it's all code based, it can handle things like retries, rate limits, pagenation, web hooks, all of that stuff that can definitely be a bit of a pain to build in manually in a workflow builder like Naden. Of course, you still need to go grab your API keys, but you won't wrestle with like low-level technical details anymore because your energy now is going towards what actually happens, not how to configure the API call. All right, so now that you know the four main changes, let's actually just go ahead and build a super quick agentic workflow together. Okay, so for the actual live build of an agentic workflow, I thought that we would do a lead generation automation because it's one of the most common types of automations that I get requests for and you're going to see how easy it is to actually spin up. So before we actually hop into Cloud Code, which is what we're going to be using today, I just wanted to familiarize everyone with what we're actually going to be looking at because otherwise I know it seems a little intimidating. So basically, we're going to be in VS Code using Cloud Code, but we're not going to be coding at all, right? We're going to be using natural language talking to an agent that's going to live on the main section of the screen. And this is where we're going to be able to do our planning. It's going to ask us questions and it will start executing, you know, writing code and taking action. On the lefth hand side is where we're going to see essentially our project. So inside of our project, we're going to have different files. We're going to have a system prompt. We're going to have different workflows. We're going to have different tools. So very similar to like what you were used to, but it just looks a little bit different. In order to do all this, we're going to be using a framework called the WAT framework, which basically just stands for workflows, agent, and tools. So in formal software engineering or different types of vibe coding, you may hear it be called different things, but essentially it's the same concept. Workflows, agent, tools. The agent is the brain that we talk to that helps us figure out what workflows are we going to be looking at. So in this case, the workflow is lead generation, but within that lead generation workflow, there's lots of different actual actions and tools that the agent needs to use in order to scrape the leads, put them into a Google sheet, do analysis, things like that. So just like in any how you have your workflows, there's tons of different actions and tools within that workflow that make it up. So, same thing here. It just looks a little bit different, like I said. All right. So, now here we are in VS Code. This is the agent. You can see I'm having a conversation with it here. And on the lefth hand side is our files. I know there's a few more than what I had in the Excal, but really the key ones to look at are tools folder, our workflows folder, and then our claw.md file. And you can see right here, it already set up our project. And this is what we have inside of the structure. We've gotten files. I'm not going to explain that right now. We've got tools, which will be our Python scripts for deterministic execution. We've got our workflows which will be markdown SOPs defining what to do and how. We've got env.get ignore. Just ignore that stuff for now. And then we've got claw.md which are our agent instructions. So basically the system prompt for this whole project. What are we doing here? How do the folders work? How do you execute actions? So like I said, this section is not going to be a full tutorial. I just wanted to familiarize us all so that we can see what's actually going on. So down here you can see that I just put the Claude agent on plan mode so that I can actually talk to it and have it ask us questions so we can make really really clear instructions before we actually start automating anything. Hey Claude, I need you to help me build a lead generation automation. I want to scrape tons of different dentists in Chicago, Illinois so that I can reach out to them and ask them if they want my AI automation services. My agency is called UpAI. We provide custom AI consulting and implementation. and I need you to do research about these people, scrape the leads, write them a personalized outreach message, and then put all of this data in a Google sheet for me so that I can just start going to work. So that request that I just shot off is something that 100% of you guys could do and probably want to do. So now that I've shot that off, basically what we're seeing here is we are watching Claude think about what to do. So it's going to look through our files. It's going to see what exists already just to familiarize itself with its environment. And then it's going to start creating a plan. And as it's creating this plan, once again, we can watch and look at everything that it's doing. But what you're going to see is it's going to come back to us with some questions. And then once we provide those answers, we can actually just let it start going. So now we're at the question phase. You can see it just said, "I have a draft plan, but I need to ask you some questions." So first it says, "What data source should we use for scraping dentist leads?" I'm going to go ahead and just use the recommended Google Places API. For enrichment, it says, "How deep should we go?" Let's just go with basic. For message tone, we can be professional, friendly, or premium. Let's just go ahead and be friendly. And then for API setup, it says, "Do you already have the required API keys set up?" So it'll either help you get that all set up, or you can say, "Yes, I have them." them. So, I'm just going to go ahead and say yes. And once I submit those answers, it's right back to work. It's going to keep working on the plan. If it has any more questions, it will go ahead and ask us. Okay. So, now we have our full plan done. Chicago dentist for UPAI. We've got an objective. We have our preferences. It goes through the full architecture as you can see. So, it's basically going to create a Chicago dentist leads workflow and that will be shown over here in the workflows folder. It's going to create three different tools. So, one to scrape dentists, one to generate outreach, and one to export to sheets. And those will all show up in the tools folder. Over here, you've got implementation steps. It's going to add our API keys in AENV. It's going to build the tools and then it's just going to keep going on. And obviously what you would do is you'd read all this and you would say yes, I accept or I'm going to keep planning and things like that. But we're just going to go ahead and auto accept what it came up with. And so when it starts to implement, it builds a to-do list. As you can see, it's got these five tasks. And then we're going to go ahead and be able to watch it execute across these to-do tasks. At this point, you can see it's created the files we need. So it went ahead and created three different tools. Export to sheets, generate outreach, and scrape dentist. You can see all of these are actual Python scripts that it wrote in code and then it created the workflow file which is Chicago dentist leads. So now anytime that I ask this claude agent to find me some Chicago dentist leads it will basically come reference this workflow understand what it needs to do and how it works and then it will just go do it for me. So the last thing that I need to do here is give it my actual API keys so that it can use Google places, use OpenAI and then use my Google Sheets. So just like that in a matter of minutes I have an automation built here in cloud code that I can continuously hit and get as many leads as I want. If we go look at the actual output that it gave us when it was just testing out 10 leads to see if it worked, we can see that each of these have a name, an address, a phone number, a website. They've got rating, reviews, neighborhood, subject line, and then we have our actual outreach message, which is personalized for them and for us. You can see that they mention things like the fivestar rating or the 4.9 star rating from 714 reviews. And so obviously, this was basically very minimal prompting. So what we would do now is we would come back with natural language and say, "Okay, you know what? I don't actually want to limit it to just Chicago. Let's also search in California." Or we' say, "The outreach messages aren't super personalized." Or you could say, "The outreach messages aren't really custom enough. Let me drop in a PDF of all the information about my business up at AI and incorporate that in all of your outreach messages." You could say, "Hey, I'm realizing that we don't have any emails. Could you actually incorporate emails into this automation as well?" And so all of those changes that I ask it to make, it would actually go ahead and it would change the workflow file. It would change the tool files so that next time it does all of this automatically. And so of course you can see already there's a lot of power in just being productive by making Claude Code essentially an executive assistant having different workflows and different tools for everything that you do on a daily basis. You can go ahead and trigger Claude code to do work on something else on your other monitor and come back and it will be done for you. But if you need to have an automation that triggers based on some sort of event or goes off once a day. You don't have to be here watching Cloud Code to make it go off. You could of course publish these scripts and the workflows onto something like Modal in order for them to actually run without you being around at all. So, I know that I went pretty fast through this demo, but that's exactly what this was, just a demo. I'm going to obviously be making full tutorials and deep dives on how you can actually do all of this stuff yourself. But really, I just wanted to give you a taste of what this is like and show you in a matter of minutes what I was able to do in here, which would have taken me at least an hour or so in eden. And remember that I didn't have to set up any sort of API calls or look through any documentation on my own. Cloud code handled all of that messiness. So, now that we've talked about those big shifts and you've seen me build one right in front of you guys, let's talk about what the future's going to look like with this stuff. So, I've got four main topics to hit on here. Fully autonomous workflows, agents managing other agents, protocols like A2A, and then longunning project agents. So, starting off here with fully autonomous workflows. Right now, most workflows are reactive. They wait for something to happen. So, their trigger can be like a web hook, a schedule trigger, a form submission, and then they go ahead and they take the steps that are in the workflow. But the next generation of these agentic systems, they're not going to wait around. They're going to be proactive. So, they're going to constantly scan your tools like your CRM, your inbox, your project management software, and they're going to look for inefficiencies or roadblocks or risks. They'll spot things like deals going cold before you even notice. They can flag projects that are falling behind before anyone misses a deadline. And they're not just going to alert you, but they might actually propose fixes or even just take action on their own. So, Deote predicts that 25% of enterprises using generative AI will deploy agentic pilots this year, jumping to 50% by 2027. By 2028, agents will act as autonomous partners handling complex multi-step problems and proactively influencing decisions. So, we're talking about proactive teammates that look ahead and say, "Hey, here's what stuck and here's what I already did to help you unblock it." And the market is starting to reflect that. The AI agent market is projected to grow from roughly 8 billion in 2025 to high 40 or low 50 billion range by 2030, which is a 43% compound annual growth rate, which is of course driven by workflow automation use cases. Next, agents managing other agents. So, instead of just one giant all-purpose agent, we're moving towards teams of specialized agents. An email agent, a research agent, a reporting agent, a data cleanup agent, and a manager agent that can delegate the work, review output, and stitch everything together. Think of it like a project manager with a remote team, spinning up the right specialist to fix problems without waiting for you to kick anything off. Now, these multi-agent systems are under development with the first pilot starting in late 2024 and probably even earlier. And I know this is stuff that's kind of been around already for a while, but research is showing that multi-agent setups often outperform a single model by distributing tasks among specialists. So big companies are starting to prepare for this pattern. So it's not just one bot per workflow, but it's agent teams embedded across, you know, sales, support, operations, finance, and they'll coordinate with each other. Like I said, we've been seeing this in NN, you know, sub agents or ultimate assistants that manage different agents. And we can see that we're getting better results when each agent has a specialized task. But that was just a preview of what's to come with agents managing teams of agents. All right. Third, we've got protocols like ATA. So MCP is how agents talk to tools. A toa or agent to agent is how agents talk to other agents. In April of 2025, Google Cloud announced A2A as an open standard so that AI agents from different vendors can communicate, share context, and coordinate across, you know, different systems. And that was huge because we know in business processes, it's rare that like one department never has to talk to multiple other departments in the process to get certain things done. So this protocol defines agent carts basically describing what each agent can do, a shared life cycle of tasks and how agents share context securely. So one agent can say something like find me candidates, another handle scheduling. Another one does background checks and they coordinate without you playing middleman between all of them. So Google launched the ATA with support from firms like Salesforce, SAP, Service Now, Workday, and over 50 other enterprise partners. And that level of coordination around a standard is a loud signal. We're definitely heading toward an agent mesh where everything plugs into everything. And last but not least, we've got longunning project agents. So, this is the one that I'm really excited about because right now agents are really good sprinters, but they're not good long-distance runners because they can crush a one-off task or deliverable, but if you ask them to manage a project for weeks, then they start to forget things, repeat themselves, hallucinate, drift off the rails. And we actually have benchmarks proving this. So, there's a vending bench, which basically asks LLM to run a vending machine business over a long simulated period. And even the really strong reasoning models show high variance and they get worse over time. They do things like forgetting orders, mistracking inventory, and falling into repetitive loops. So, why do I think the future is different? Because we're starting to see techniques that solve this. So, one of those is continuous loops. So, there's a plugin called Ralph Wiggum for Claude Code that keeps looping tasks back to the agent until certain success conditions are met. It adds guard rails like max iterations and explicit done signals. So, it's not perfect by any means because there's still a lot of elements we need like human in the loop and midexecution course correction, but it's moving in the right direction and it's evolving really fast. We've also got stuff like agent harnesses with shift based work. So, Enthropic is developing effective harnesses for longrunning agents. One agent works for a while then leaves structured artifacts like notes, to-dos, differences, what changed, what broke, what's next. Then the next shift picks up from there and looks at all that kind of stuff. So, instead of keeping everything in a huge giant context window, which is where we start to see issues, it has more of a notebook and a checklist. So, if you think about it like an actual human, if it was to sit there and code for, you know, 72 hours straight, the human would get tired and the work would get sloppy. But if the human was able to just work for 10 hours and then another one takes over after a fresh night's sleep, it can read all of the documentation about what's happened and what it needs to do and then that human just picks up on a new shift. Now, we're not fully there yet, of course. It's not perfect, but the fact that we have benchmarks that are measuring this and we have, you know, longunning harnesses being developed and techniques like the Ralph Wigum loop, this tells you exactly where the industry experts want this to go and expect it to go. agents that stick with a goal for days or weeks, maybe even months, and they keep track of what they've done and they keep improving your systems while you sleep. So, here's the bottom line. Analysts and experts are saying that within a few years, half of companies using generative AI will have deployed agentic systems. Google and others are already standardizing how agents talk to tools and to each other. If you've been building an end to end and you understand how to map out processes, you're exactly the person who can step into this next layer because again, the skill isn't coding. The skill is designing what agents should actually do, where they should be proactive, and how they should all work together. That's where the value is going. And that brings me to what's going to be changing with this channel. So you're going to be seeing me doing a lot more audentic workflows. So diving into things like cloud code, anti-gravity, and more IDE based tools as they come out. But my mission hasn't changed. I'm here to bring you the most valuable skills to learn and practice how to actually make money, improve a business, improve your own business. Now, you guys know I don't come from a technical background, and I know that most of you don't either. So I'm never going to dive into something if I don't think that I can explain it clearly and I'm never going to cover anything if I don't think it's worth your time learning. That's my promise. None of us know exactly where the space is going, but I promise that I will stay as close to it as possible and bring you what's most important because there is a lot of noise and a lot of changes out there. And I know it may seem right now intimidating to look at something like a cloud code environment, but I also know that the IDEs are going to get simpler and easier to use over time as well. Now, I know that some of you might be thinking, does this mean that all of the time I just spent learning noden was a waste? No, do not think like that. The opposite, actually. If you've been building an end, you're ahead of the curve. And I've got some reasons why. The first one is process decomposition. You've already learned how to break business processes into discrete steps, handle edge cases, think about what happens when things fail. And that's exactly what agentic systems need from you. The next one is because you have systems vocabulary, so you now actually know what's a web hook, what are triggers, what's API authentication, data transformation, conditional logic, so that once again, when you're directing an agent, you can speak precisely about what you want. If someone doesn't have that foundation and they jump straight into Agentic Workflows, they'll be able to get stuff to work, but they probably won't be able to prompt it as efficiently and they'll just say things like make it work with my CRM rather than you being able to say trigger this on the deal stage change, then pull the contact object and transform these fields, make a post request to my endpoints, blah blah blah, which in turn will have the agents building things faster and better because your instructions were better. You also have some more intuition about failure. So, you've seen so many things break in. who understand how debugging works, which means you know the patterns that typically cause errors like rate limits or malformed JSON or token expiration or weird edge cases that you may not have thought of if you hadn't built automations in NN already. So I've obviously poured in like the past full year of my life into Nodn but Naden wasn't a detour. It was something that I wanted to go through because now like I think about that like learning to ride a bike. I've got balance. I know how to steer. I know how to pedal. But now we're going to be hopping on to like a motorcycle where we don't have to pedal necessarily but we still have to learn how to balance and steer. So yes, the barrier to entry to building all of these agents and AI systems is dropping. But what businesses actually pay for is not just the actual build itself. They want someone who can understand their problem, which most clients can't always articulate what they need. They need someone who can integrate with their existing mess. Legacy systems, weird edge cases, compliance requirements, things like that. And of course, someone who can keep optimize, iterating based on real usage, expanding the scope, maintaining systems over time. So yeah, the implementation layer is getting easier. Everything else that you do after that is getting way more valuable because more people will attempt automation and fail at the parts that aren't implementation. User tooling expands the market, which means more businesses can actually access the stuff because it's not as expensive as it used to be. But they'll still need someone who knows how it works and knows how to use it. So your job essentially shifts from just being a builder to more of an architect and a manager and once again, like I've been saying, a consultant. So if you want to stay tapped into all of this, then check out my free community of over 230,000 members. I'm always dropping free resources and templates in there. And if you're looking to dive deeper, then I've got a paid community of over 3,000 members who are building businesses around this kind of stuff every single day. And that's where I'll be diving a lot deeper into this kind of stuff, too. So, the link for those are in the description. But that's going to do it for this one. I know that this was a ton of information. So, I threw all of this into a resource guide, which you can access for completely free in my free school community. And if you enjoyed or you learned something new, please give it a like. It definitely helps me out a ton. And as always, I appreciate you guys making it to the end of the video. I'll see you on the next one. Thanks everyone.