Learning Library

← Back to Library

Avoiding the n8n AI Agent Trap

Key Points

  • The speaker addresses a common frustration: non‑technical users want to build custom AI agents without deep coding, finding tools like LangChain too complex and out‑of‑the‑box platforms too limiting.
  • While visual workflow tools such as N8N (referred to as “NAD”) empower creators by democratizing automation, that same flexibility often becomes a “complexity trap” that leads to tangled, hard‑to‑maintain agent implementations.
  • Real‑world examples show organizations ending up with hundreds of poorly maintained agents, high costs, and dependence on the original builder, highlighting the need for scalable best‑practice patterns.
  • The video promises a “Goldilocks” approach—providing a middle ground of customizable yet manageable agents that non‑programmers can build reliably without sacrificing maintainability.

Sections

Full Transcript

# Avoiding the n8n AI Agent Trap **Source:** [https://www.youtube.com/watch?v=zRr24Mku3r4](https://www.youtube.com/watch?v=zRr24Mku3r4) **Duration:** 00:24:12 ## Summary - The speaker addresses a common frustration: non‑technical users want to build custom AI agents without deep coding, finding tools like LangChain too complex and out‑of‑the‑box platforms too limiting. - While visual workflow tools such as N8N (referred to as “NAD”) empower creators by democratizing automation, that same flexibility often becomes a “complexity trap” that leads to tangled, hard‑to‑maintain agent implementations. - Real‑world examples show organizations ending up with hundreds of poorly maintained agents, high costs, and dependence on the original builder, highlighting the need for scalable best‑practice patterns. - The video promises a “Goldilocks” approach—providing a middle ground of customizable yet manageable agents that non‑programmers can build reliably without sacrificing maintainability. ## Sections - [00:00:00](https://www.youtube.com/watch?v=zRr24Mku3r4&t=0s) **The Hidden Cost of DIY AI Agents** - The speaker warns that beginners who gravitate toward highly composable low‑code platforms (like N8N) can quickly amass dozens of fragile, costly agents that break, become unmaintained, and drain resources. - [00:03:58](https://www.youtube.com/watch?v=zRr24Mku3r4&t=238s) **Simplifying Enterprise Workflows with JSON** - The speaker argues that adopting JSON‑based workflow representations and treating the platform as a collaborative hub can make automation reliable, simple, and scalable, citing StepStone’s 25× speedup in API integration as proof. - [00:07:31](https://www.youtube.com/watch?v=zRr24Mku3r4&t=451s) **Simplicity, Refactoring, and JSON Workflows** - The speaker emphasizes that engineers refactor code to keep systems simple, maintainable, and scalable, and recommends defining workflows in JSON and using LLM‑generated documentation to preserve clarity and ease of future maintenance. - [00:10:47](https://www.youtube.com/watch?v=zRr24Mku3r4&t=647s) **Microservices: Simplicity Through Separation** - The speaker stresses that true simplicity relies on composable, separated concerns, recounting Amazon’s transition from a massive monolithic codebase to microservices as a practical illustration of why modular architecture is essential for maintainable, scalable software. - [00:15:01](https://www.youtube.com/watch?v=zRr24Mku3r4&t=901s) **Siloed Automation Undermines Teamwork** - While personal automation can save massive time and streamline predictable workflows, treating it as an individual, undocumented tool creates maintenance, debugging, and scalability issues that jeopardize team productivity and continuity. - [00:18:14](https://www.youtube.com/watch?v=zRr24Mku3r4&t=1094s) **High‑Bar Automation with LLM‑Powered NAD** - Leaders must demand solid engineering fundamentals for marketer‑built automations, while LLMs serve as accelerants that let non‑developers embed AI‑driven workflows into real business tasks through NAD. - [00:21:25](https://www.youtube.com/watch?v=zRr24Mku3r4&t=1285s) **Balancing Accessibility and Discipline in AI Agent Development** - The speaker warns that while tools like Naden make building AI agents easy, developers must still apply solid engineering principles—simplicity, separation of concerns, maintainability, and proper documentation—to achieve real ROI and avoid fragile, burnout‑inducing solutions. ## Full Transcript
0:00Today I want to give you a key to one of 0:03the most persistent questions that I get 0:06in the AI space. It's about AI agents 0:08and it's specifically I need to get 0:11started building AI agents. I'm not 0:13sophisticated enough. I don't have 0:14enough code to go and do it with just 0:17code. So they're not going to use 0:18Langchain. They're not going to use Lang 0:20Smith. They want simple agents, agents 0:24they can use, and they want custom 0:26agents. And so a tool like Lindy.ai AI 0:28that has a lot of outof-box agents and 0:30some composability feels limiting to a 0:33lot of the people I talk to in this 0:34situation and a tool like feels just 0:38about right because it's more 0:39composable, more customizable. You can 0:41bring some code and so they feel like 0:43they're in a playground. But if you are 0:45getting started with agents, you need to 0:47recognize that that composability, that 0:50configurability, the power you feel with 0:52N8N is the trap. That is the trap. And 0:55that is why this video exists because 0:58people get started with something like 0:59NAND and they get so excited and the 1:02stars come out in their eyes and they 1:03feel the AGI, they feel the AI agents. 1:06They know they can do all this cool 1:08stuff and then they find out that the AI 1:11agent breaks. They find out they have 1:13556 of them across the business and God 1:16knows what's happened to 332 of them and 1:19only 50 of them are being used regularly 1:21and it's all adding up to a big bill. 1:23They find out the person that built the 1:25agent went on vacation and now they're 1:26they don't know how to fix it because 1:28this this like drag and drop thing has 1:30created a tangled mess and it's 1:31impossible to see. That is all real life 1:34examples. I have seen it over and over 1:37again. This video is going to help you 1:39get from I want to do custom agents to I 1:43can do custom agents. I can use Naden. 1:46It's a great tool, but I know how to I 1:48know how to do it well. I know how to 1:50follow best practices when building 1:52agents and I don't have to be a custom 1:55coder to do it. I have written 1:57comprehensive guides to agents. This is 2:00more narrowly focused. It's very 2:02comprehensive, but it's more narrowly 2:03focused on what I call the Goldilocks 2:06use case. It's not I am so out of box. I 2:10just want a pre-built agent for my 2:11calendar. It's also not I am so 2:14freestyle I'm a developer and I can do 2:16anything. It's in between. It's the 2:17people who want customizability without 2:20committing to just a code lifestyle. So 2:22the first truth to understand if you're 2:24in this bucket is that NADN has a visual 2:27workflow builder that genuinely 2:29democratizes automation and that's part 2:31of why people gravitate to it. For the 2:34first time it is really genuinely 2:36possible and this does work for 2:38non-programmers to build sophisticated 2:40AI agents just by dragging and dropping 2:42little nodes into a little tree. The 2:44second truth, which I've already hinted 2:46at, is that the same visual builder is a 2:49complexity trap. It has killed so many 2:52NAD implementations. So, ironically, the 2:55very feature that makes you want to use 2:57it is the feature that becomes 2:59unmaintainable at scale. But, but worry 3:01not, I'm going to give you a way 3:03through. I want to map out how to get 3:06from that honeymoon phase where you 3:07install an you you you start to build 3:10your first node, you start to see data 3:12flowing across the screen. It feels so 3:14cool. It works. You feel like you've 3:16unlocked superpowers. You want to build 3:1810 more of them. I want to get you 3:20accelerated through the trough of 3:23disillusionment where you have to add 3:25error handling. You have to add 3:27conditional logic. That's even more 3:28nodes. The edge cases pile up. You tell 3:31yourself it's manageable, but the clean 3:32workflow looks like, I don't know, the 3:34subway map of Manhattan. It looks like 3:36like ridiculous, like a pile of 3:37spaghetti. At a certain point, if you 3:39are serious and you are only using the 3:41nodes and dragging and dropping, you are 3:43going to get to a point where you have 3:4512 workflows and six 633 nodes and it 3:48fails at 2 a.m. and you're spending 3 3:50hours debugging it and you feel like you 3:53are in hell. It is an incredible pain. 3:56You can't simulate inputs correctly. 3:58There's nodes that have failed. There's 4:00an LLM call that's different because 4:01LLMs have updated. You have error 4:03objects when they fail and you don't 4:05know what that means. This is where you 4:08can actually shift your approach. You 4:11don't have to go down that path. You can 4:13actually build agents that are useful. 4:15As an example, a real company, 4:17StepStone, it runs 200 mission critical 4:19workflows in NADN. They achieved 4:22approximately a 25x speed up in API 4:25integration time. The 200 workflows that 4:28takes focus. If you are going to do it 4:30at the scale of a company, you almost 4:32need to treat NN like a hub of 4:35excellence where your product and your 4:37marketing and your CS teams meet your 4:39technical teams and everybody can focus 4:42on creating reliable, simple, clear, 4:46followable workflows. And that's 4:47something I'm going to emphasize again 4:49and again and again. When you are 4:51building, make sure they're reliable, 4:52simple, and clear. And you might say, 4:54Nate, is this whole video about you 4:56telling me that my workflows in NAD are 4:58complicated? I already knew that. No, 5:01this is about suggesting to you that you 5:04can get farther with your workflows if 5:07you put them into JSON representations 5:10because JSON representations tend to 5:13force simplicity because you typically 5:15build them in LLMs that tend to obsess 5:17over making workflows as simple as 5:19possible. JSON representations of 5:22workflows are like kitchen instructions. 5:24They tell the chef what to make. And one 5:27of the things that makes them powerful 5:29is that yes, you can drag and drop a 5:31workflow, but you can also do this 5:33programmatic kitchen instruction thing 5:35where you just hand the chef the JSON 5:37workflow and say do this and it will 5:39work. And I want to suggest that JSON 5:41representations are useful here. Not not 5:44because LLMs have some magic power with 5:47JSON. and I saw that circulating around 5:48the web and I kind of rolled my eyes. 5:50That's not what I'm trying to say here. 5:52It's just that JSON is a good ingest 5:54method that this particular tool NAND 5:57happens to understand that LLMs can 5:59write well and that you can use as a 6:02carrier or a vessel for the clarity of 6:04vision that you have. JSON acts as a 6:07simplifier for you. If you work with 6:08claude, if you work with chat GPT to 6:11build a JSON workflow for your N8 6:14automation, you are so much more likely 6:17to have a workflow that works, 6:20especially if you are invoking the 6:24documentation that NAD publishes about 6:27their workflows and their tool calls 6:29because then the LLM will know that when 6:31it's writing the workflow and it will be 6:33less likely to have errors. I also want 6:35to suggest to you that if you are in the 6:37business of building agents, it is 6:39probably better to understand that you 6:42are in the business of building software 6:44even if you're not a developer. I don't 6:46want that to scare you, but I try and 6:48convey it honestly because I don't want 6:50people to be surprised. Effectively, 6:52instead of programming, you're 6:54configuring and so it feels easier, but 6:56it's actually going to be more 6:58strategic. It's going to require more 7:00thinking and intent on your part. And 7:02it's going to require you to understand 7:04some things that most software engineers 7:07have built into their DNA that doesn't 7:09come built into other job families. I am 7:12here to bridge that gap for you so your 7:14agents work better. Let's start with a 7:17simplicity principle. You need to build 7:20workflows that are as ruthlessly simple 7:24as they can possibly be in order to 7:27maintain them. Simple, simple, simple. 7:29Drill it into your head. The reason why 7:31is that simple is maintainable. This is 7:34why engineers will tell you they need to 7:36refactor the back end of the codebase. 7:38And you wonder what they're doing. Well, 7:40let me tell you, engineers out here are 7:42nodding their heads. They know what 7:43they're doing. They have to refactor the 7:45back end of the codebase because they 7:47have to make it simple and maintainable 7:49and scalable. Simple is scalable. Simple 7:51is maintainable. Simple is readable. 7:53Effectively what you have with NAND and 7:56agents is you have a combination of 7:59function and documentation in one 8:01format. That spaghetti code that looks 8:03like the subway map of Manhattan that is 8:06both the actual diagram of the workflow 8:09and also your only documentation. That 8:12is why it hurts so bad. The advantage of 8:15composing these workflows in JSON is you 8:17can not only be more accurate and be 8:20more likely to be simple. If you're 8:21working with an LLM that has a bias to 8:23simplicity, you can tell it to be 8:24simple, right? It will help you get 8:26there. You can also use that exact same 8:30LLM to write the documentation for the 8:33JSON that you're giving. Hey, this is 8:35the documentation so I can save it and I 8:37can maintain it and I can know why I 8:38made the decisions I did. I would 8:40encourage you to do that because you 8:42want to be in a place where you're not 8:43the only one maintaining it unless 8:45you're a solo builder, which by the way, 8:47it is possible to have a solo business 8:49that works this way. As an example of a 8:51small business that did this, uh, a 8:53company called Border, B O R DR, they 8:55dropped the E, built a real business 8:57helping people navigate Portuguese 8:59bureaucracy. This is for part of the 9:01sort of expatriate workers movement. You 9:04can work around the world, the nomad 9:05movement, etc. They used NAND to do it. 9:08The entire operation runs on NADN 9:12workflows. I want you to guess how many 9:14of those little boxes on NADN they have 9:16for their core workflows. Oh, go ahead. 9:18Guess. All right, you guessed. Great. 9:20They have 18. Not 180, not 18,000. 18. 9:25They understand that complexity 9:27compounds risk. Complexity also 9:29compounds exponentially in automation. 9:32This is just basic graph theory. If 9:34you're a developer, every node you add 9:36does not just add one more thing to 9:38maintain. It adds interactions with so 9:41many other nodes. A 10 node workflow has 9:44something like 45 possible interaction 9:46points. A 20 node workflow, just 10 9:48more, would have 190. A 50 node workflow 9:51would have over 1,200. Now, Portuguese 9:54bureaucracy is legendarily complex, 9:57which is why the business exists. Their 9:59workflows are simple not because the 10:01problem is simple but because they 10:03understood how to decompose complicated 10:07problems into composable parts. Every 10:10workflow does one thing. It does it well 10:12and it can be understood quickly. That 10:14is a principle of software engineering 10:16that I don't think is widely enough 10:18understood by people getting started 10:20with NAND and building agents. people in 10:22that Goldilocks bucket, they almost all 10:25overwhelmingly tend to come from the 10:28non-technical side. And by the way, if 10:30you're an engineer listening to this, 10:32this video is kind of your get out of 10:34jail free card when the director of 10:35marketing draws this ridiculously 10:37complex like whiteboard full of nodes 10:39and says, "This is what we want for our 10:41agent." And you're just kind of 10:43screaming on the inside. Show them this. 10:45The simplicity principle matters. It 10:47really does. And by the way, I said 10:50composable parts on purpose. One of the 10:52fundamental aspects of software 10:55engineering is the idea of separation of 10:58concerns. Another way to put it is 11:00things go in their proper place. You 11:02keep your concerns neat and tidy so that 11:05you can manage them well later. This is 11:07where at a larger scale we get the idea 11:10of microervices. 11:12We're gonna take a tiny diversion here 11:14because I worked at Amazon and Amazon is 11:15where the idea of microservices and APIs 11:19originated. Yes, we invented them along 11:21with lots of other things. I don't know 11:22if you consider that a curse or a 11:23blessing, but you've got it anyway. 11:25Microservices are the idea that the 11:28separation of concerns is so important 11:31in software that you cannot scale at a 11:34certain point unless you have it. We 11:36started off in coding by building 11:38monoliths. What what we would call the 11:41subway map of Manhattan for NADN. That 11:44was software. It was software. Even at 11:46Amazon 15, 20 years ago, everything was 11:48this one giant codebase. Well, that 11:50makes it really hard to maintain as you 11:51get bigger and bigger and bigger code 11:53bases. You don't know what a piece of 11:55code touches. Only the senior engineer 11:58who never takes a vacation knows. And 11:59God help him. I hope he doesn't get hit 12:01by a bus. Well, microservices existed 12:04because we had to have them exist to 12:06actually build. We had to separate 12:08concerns. And now every engineer knows 12:11about it because it's just such an 12:12obviously good idea. You separate out 12:15the components of the software. You 12:17standardize how those components 12:18exchange data and business rules. That's 12:20what we call APIs. And then you have 12:23separate concerns you can manage 12:24cleanly. Apply that principle when you 12:28are building agents. Don't build a 12:30monolith agent that does everything. 12:32Separate out your concerns cleanly. 12:35Separate it out cleanly so that you only 12:37have to do one or two things in a 12:39particular agent. It's really, really 12:41critical to focus. I want to give you 12:42another example. This is also a real 12:44example. Delivery Hero save hundreds of 12:47hours a month with their NAD automation. 12:49They handle hundreds of requests 12:50automatically. And you notice what they 12:53automated? IT account recovery. Not the 12:55whole IT department, not employee 12:57requests, just one well-defined process. 13:00This is reinforcing what I'm saying. If 13:02you want to be successful in an agent 13:04implementation, you must focus 13:07radically. Identify one process. It 13:10needs to be painful. It needs to be 13:12frequent. It needs to be really 13:13well-defined with good edges. Automate 13:16it all the way. Run it. Obsess over it. 13:19Learn what breaks. Fix the breaks. Only 13:22when it's mature, sustainable, well 13:25doumented, do you move to the next 13:27process. The typical failure pattern I 13:29see is the opposite. Teams don't clearly 13:32define the pain. They don't know where 13:34the edges are. They had a seminar and 13:37the seminar said use AI agents and they 13:39got all excited and they try to automate 13:42everything at once. They build giant 13:44sprawling workflows and it touches 13:45multiple systems and it looks incredible 13:47and on day one it all works and they 13:49have a CEO announcement and they're 13:50creating dependencies they don't 13:52understand. And when something breaks 13:53and it always breaks, plan for it to 13:56break. They can't isolate the problem 13:57because nobody can read the map anymore 13:59because everyone's on to the next 14:00project because the CEO declared victory 14:02on AI agents. AI agents are not a 14:05tickbox. AI agents if you want to 14:08implement them implement them this way 14:10and so many teams do. AI agents are just 14:13a new way of doing software for 14:16everybody. And it's possible to do 14:18software for everybody. I firmly believe 14:21that if you work with your favorite LLM, 14:23whether that's Claude or Chad GPT or 14:25Gemini or even Grock, you can get to a 14:28point where you build this yourself and 14:30it builds in line with what I'm 14:31suggesting. It's simple. It's 14:32composable. It fixes real pain points, 14:34etc. You don't have to have an engineer 14:36to do that, but let's recognize that you 14:39are doing real engineering work. you are 14:41actually building software. And this is 14:44part of the magic of the AI era. If 14:46you're willing to think this way with 14:48this kind of clear intent, you too can 14:51do the work of building software and you 14:54can generate the magic, the real ROI 14:57that comes from true automation. 14:59Remember when I said delivery hero saves 15:01200 hours a month? They do. It's real. 15:03Because it turns out at scale a lot of 15:06people need their IT accounts recovered 15:07and that's a very predictable workflow 15:09and you can recover it. Border can 15:11maintain their scale and navigate 15:14Portuguese bureaucracy because they 15:15built these agents effectively and it's 15:18not the tool that's the difference. 15:19These guys are using N82. They just know 15:21how to use it. So if you're in this 15:23Goldilocks use case, you should be 15:25encouraged that yes, this is a real use 15:29case. It really matters. You really can 15:33automate it. You don't need to wait for 15:34the developers, but you got to take it 15:36seriously. One thing that I want to 15:38emphasize at this point, we've talked 15:40about best practices for coding, best 15:42practices for software, the team 15:44problem. Your private automation is not 15:48a team level product. Nobody talks about 15:51this. Nad likes to market this as 15:53individual productivity because frankly 15:55it gets them more seats and customers. 15:57You can build an agent that works 15:58perfectly for you. You might understand 16:00its quirks. You know how to restart it 16:01when it hangs. You remember to clear the 16:03memory cache weekly. You know how it 16:04fails on PDFs over 10 megabytes in size. 16:07And then you go on vacation. And it 16:09turns out this was producing a 16:11deliverable your team cared about. Your 16:13team can't debug your workflows when 16:15they break. They don't know why certain 16:16design decisions were made. They're 16:18afraid to modify anything because they 16:19might break something else. It looks 16:21like this scared spaghetti on the wall. 16:23This is how automation projects die. 16:26They die not really from technical 16:29failure. They die from knowledge 16:31isolation, from silos. Successful 16:34transitions require documentation that's 16:36useful. Remember when I said document, I 16:39said document because your automation is 16:41your team's problem. Don't create huge 16:44manuals that nobody reads. Just create 16:47very simple, very short runbooks. When 16:49this error appears, check this. This 16:51workflow depends on this web hook. Clear 16:53this cache every Monday or response 16:55times degrade. Patterns, patterns, 16:57patterns. You don't need to be creative. 17:00You need to have clean patterns. Every 17:02workflow ideally would follow the same 17:04error handling pattern. Every agent, if 17:06possible, should use the same memory 17:07config. Is it boring? Yes. Is it 17:10maintainable? Exactly. That is the 17:13point. The more we make NADN and agent 17:17automations a teamle product, the better 17:20off we are going to be. And this by the 17:22way is one of the hidden lynch pins in 17:25AI strategy and AI work. We typically 17:29talk to the seauite and we talk to 17:31individual contributors and senior 17:33managers and directors people who run 17:35teams are left out of most of the 17:36conversations. We either when we're 17:38talking to IC's implicitly assume 17:40they're the leadership with the seuite 17:42or if we're talking to leadership we 17:44implicitly assume they're with the IC's. 17:45They're neither and they are critical to 17:48the AI revolution and this kind of 17:50example shows why. You cannot make an N8 17:54automation a seuite problem. It's way 17:56too high level. But you also cannot make 17:59it an individual contributor problem 18:01because that leads to all kinds of 18:03downstream breakages and workflows as 18:05I'm hoping I am calling out here. It's a 18:07team problem which means it's a director 18:09problem. It's a senior manager problem. 18:11You guys in those positions need to be 18:14the ones that are insisting on a high 18:16bar here. insisting that your team, even 18:19if they're marketers, learns enough of 18:22the basics of good engineering 18:24principles that when you touch 18:25automations and build agents because the 18:27seauite said you needed to do it, you do 18:29it well and you don't give yourself pain 18:31down the road because you can make cheap 18:33automations that tick that box and you 18:35will give yourself a world of suffering 18:37down the So, let's get to one last piece 18:40here. LLMs are an accelerant in multiple 18:43directions for NAD. One, LLM enable you 18:46to do more. They tie intelligence 18:48directly into workflows. For a lot of 18:50people who are not developers, NADN 18:52represents the first and most accessible 18:55front in the integration pathway. You 18:58can actually tie your chatbot into real 19:02work. You don't have to use the API 19:05necessarily because effectively you're 19:06proxying for that through NAD. It's you 19:09can just pretend it's not there, right? 19:11You can just go on with your business. 19:12And that is huge because people want to 19:15get real work done. They want to monitor 19:17Slack channels for customer complaints. 19:19They want to categorize sentiment 19:21analysis by urgency. They want to create 19:24tickets in Jira. They want to process 19:26records in a certain way. They want to 19:28send a daily summary to a support lead. 19:30This is all real work you can't get done 19:32in the chatbot. Claude or Chat GPT or 19:35other LLMs can generate not just the 19:38JSON config for those workflows but also 19:41the design decisions and why they were 19:44made and the documentation. In other 19:46words, you access the intelligence 19:48inside the workflow and that's an 19:50accelerator and working with the chatbot 19:53is an accelerator as well. It's a second 19:55way LLM accelerate you. What I am 19:58talking about, the reason I am making 19:59this video, this would not have been 20:01possible even 8 months ago. NAND wasn't 20:05mature enough. Chat bots weren't mature 20:07enough. They weren't good enough at 20:08checking documentation reliably. We are 20:10at a point now where it's mature. Nad is 20:13ready. You can get LLMs that reliably 20:15pull up the correct documentation. 20:16They're going to give you reliable 20:18workflows. They're smart enough to give 20:19you good documentation and really help 20:21with constructing agent ecosystems that 20:23work at the team level. Th this is 20:26actually not as hard as you think. You 20:28really can build a couple of simple 20:31workflows, monitor the heck out of them, 20:33check for errors, and go on and expand 20:35from there. You can even get to the 20:37point if you've done this for a couple 20:38of months and you know what you're doing 20:40because you've been slow, focused. I 20:43know slow. I said slow. In this case, 20:45slow is smooth and smooth is fast. 20:47Because you've focused on implementing 20:49smoothly and only doing one edge case, 20:52you will quickly get to the point where 20:54you can do stuff that's more 20:56interesting, where it may make sense 20:58given the problem because it's well- 20:59definfined and you understand it well to 21:01use a complex memory system like 21:03retrieval log metageneration. You can do 21:04that in N82. People jump right to it. I 21:07wouldn't recommend jumping right to it. 21:08You can use multi- aent orchestration. 21:10You can use complicated tool chains. 21:12That's all there, but I wouldn't start 21:14there. Start with something simple and 21:16get to it over time. There are real 21:18savings here. Vodafone saved 2.2 million 21:22pounds with NADN workflow, right? 21:23They're big, right? That's part of how 21:25they got there. But recognize that the 21:27real ROI comes from the principles I've 21:31talked about. Naden is simultaneously 21:33one of the best, most accessible, and 21:35most dangerous tools I have seen for 21:37building AI agents. It's accessible in 21:40the sense that you can just start right 21:41now. You don't even have to listen to 21:42this video. It's dangerous in the sense 21:44that I see time after time after time in 21:46real world situations where I talk with 21:48people passionate about agents that NAND 21:50is the thing that burned them out 21:52because it was so accessible they felt 21:54like they could be non-programmers and 21:55not pay attention to good engineering 21:57principles. Please recognize that you 22:00are building real software and follow 22:03good principles. Make sure you emphasize 22:06simplicity. Separate concerns. Focus on 22:09maintainability. Focus on readability. 22:11Get your documentation done, work with 22:13an LLM, solve a problem that's well 22:15bounded. Your choice is whether to 22:18understand 22:20when each approach serves you and to 22:22have the discipline to switch between 22:24approaches when you need them. So 22:26whether you need to take a a focused 22:29approach on this particular agent or 22:32whether it makes sense to do two agents 22:33for a particular problem that requires 22:35high intention, high thought, careful 22:38architecture or you can just follow the 22:40marketing and you can just believe that 22:42you can just throw something up on a 22:43canvas and end it in and it will just 22:45work. That works great in marketing. It 22:46it demos okay. It's not going to sustain 22:49well over time. And so your choice is 22:52between building maintainable software 22:54with an approach that serves you. An 22:56approach that you have the discipline to 22:58scale over time. You remember when I 23:00said you could have multiple agents. 23:01Maybe don't start with that. But you can 23:03get there if you have the discipline to 23:05scale over time. Or you can start with 23:07what the marketing tells you and do the 23:08multi- aent and the rag and everything 23:10else right now. I understand the appeal. 23:12Agents are really cool. I like building 23:14with them to just take the time to get 23:16into agent building in a way that serves 23:19the long-term value of the business and 23:22frankly helps you sleep well at night. 23:23No one wants to get interrupted on 23:25vacation because their agent broke. If 23:27this works, it's not only going to help 23:29you work, it's going to help your team 23:32actually use agents and it's going to do 23:34something more than that. It's going to 23:36help your business understand what this 23:39agent thing is all about. And that's why 23:41I call NADN somewhat dangerous because a 23:44lot of the time what I see is that as 23:46these agents break, what businesses 23:48understand is that AI agents are fake 23:51and AI agents aren't real and AI agents 23:53aren't delivering the value that they 23:55were promised. That's not actually true, 23:57but it is true if you use them badly. 24:01NAD is like a knife. You can use a knife 24:04badly or you can use a knife well. I'm 24:06trying to give you the principles to use 24:08it well. Good luck with agent building. 24:10I know you got this.