WEBVTT 00:00:00.000 --> 00:00:03.879 Imagine you're navigating this tangled web of 00:00:03.879 --> 00:00:08.000 AI tools, feeling completely lost, beat. Or maybe 00:00:08.000 --> 00:00:11.460 you're picturing a world where really complex 00:00:11.460 --> 00:00:16.280 tasks just, they just flow effortlessly. Yeah, 00:00:16.320 --> 00:00:18.280 but what if the secret to getting that powerful 00:00:18.280 --> 00:00:20.719 AI working isn't more complexity? What if it's 00:00:20.719 --> 00:00:23.859 actually radical, almost surprising simplicity? 00:00:24.179 --> 00:00:25.760 Welcome to the Deep Dive. We're here to help 00:00:25.760 --> 00:00:27.980 transform that feeling of information overload 00:00:27.980 --> 00:00:31.280 into clear, actionable insights just for you. 00:00:31.920 --> 00:00:33.740 Today we're diving into a really fascinating 00:00:33.740 --> 00:00:36.340 concept from our source material, building highly 00:00:36.340 --> 00:00:39.840 effective AI agent teams. And here's the kicker, 00:00:39.960 --> 00:00:41.899 really. You don't need to write a single line 00:00:41.899 --> 00:00:43.740 of code to do it. Right. And this isn't just 00:00:43.740 --> 00:00:45.899 about the tech itself. It's more like a fundamental 00:00:45.899 --> 00:00:47.859 shift in how we can approach our daily work. 00:00:48.119 --> 00:00:49.700 We'll show you exactly how you can build your 00:00:49.700 --> 00:00:53.000 first three AI assistants. You'll basically master 00:00:53.000 --> 00:00:54.859 a repeatable method you can use for almost any 00:00:54.859 --> 00:00:58.119 task. We'll start by digging into the why behind 00:00:58.119 --> 00:01:00.219 this push for simplicity. Then we'll break down 00:01:00.219 --> 00:01:02.759 what actually makes an AI agent think of it like 00:01:02.759 --> 00:01:06.859 its own very precise job description. And finally, 00:01:06.879 --> 00:01:08.599 we're going to walk through a remarkably simple 00:01:08.599 --> 00:01:11.239 three -step framework so you can create your 00:01:11.239 --> 00:01:13.599 very own automated workflow. Get ready for some 00:01:13.599 --> 00:01:17.120 genuine aha moments, I think. Definitely. OK, 00:01:17.120 --> 00:01:20.159 so let's unpack this core idea first. Why does 00:01:20.159 --> 00:01:23.079 our source emphasize simplicity so much when 00:01:23.079 --> 00:01:25.879 it talks about building AI teams? It really pushes 00:01:25.879 --> 00:01:28.579 back against that common idea that powerful AI 00:01:28.579 --> 00:01:31.500 has to mean complex systems. Well, what's truly 00:01:31.500 --> 00:01:33.799 fascinating here is how the source directly connects 00:01:33.799 --> 00:01:36.560 complexity. It doesn't just say it's difficult, 00:01:36.920 --> 00:01:39.500 it links it to these deep systemic sustainability 00:01:39.500 --> 00:01:42.260 issues. The insight is that complexity isn't 00:01:42.260 --> 00:01:44.459 just inefficient, you know? It's a fundamental 00:01:44.459 --> 00:01:46.519 barrier. It stops people from adopting it, and 00:01:46.519 --> 00:01:48.819 it's like a ticking time bomb for actually using 00:01:48.819 --> 00:01:51.000 these things long term. The source actually calls 00:01:51.000 --> 00:01:53.780 complexity the enemy. Strong words. Yeah, the 00:01:53.780 --> 00:01:56.340 enemy. That's a powerful statement. And when 00:01:56.340 --> 00:01:58.700 you start to really think about it, the burdens 00:01:58.700 --> 00:02:01.299 of complexity... They become pretty clear, right? 00:02:01.379 --> 00:02:06.739 For one, these huge interwoven systems, they're 00:02:06.739 --> 00:02:10.099 just incredibly hard to maintain and update. 00:02:10.300 --> 00:02:12.419 Oh, yeah. It's like trying to fix a giant spider 00:02:12.419 --> 00:02:14.819 web, like the source says. You pull one thread 00:02:14.819 --> 00:02:17.099 and you have no idea what else might unravel. 00:02:17.280 --> 00:02:19.460 Exactly. And debugging. That must be a nightmare. 00:02:19.620 --> 00:02:22.240 Something breaks trying to find that one little 00:02:22.240 --> 00:02:24.919 faulty thread in this massive single system. 00:02:24.960 --> 00:02:27.139 It's the needle in a haystack problem. Super 00:02:27.139 --> 00:02:30.340 frustrating. Consumes so much time. So if debugging's 00:02:30.340 --> 00:02:34.280 a nightmare, then scaling sounds almost impossible. 00:02:34.560 --> 00:02:36.319 Like you're constantly playing whack -a -mole, 00:02:36.500 --> 00:02:38.199 we're adding something new, break something old. 00:02:38.300 --> 00:02:40.020 That's exactly it. And think about adoption. 00:02:40.180 --> 00:02:42.240 If it's that hard to manage, getting new people, 00:02:42.340 --> 00:02:44.900 new teams to actually use it, it's a huge uphill 00:02:44.900 --> 00:02:47.479 battle. If it's too confusing, people just... 00:02:47.530 --> 00:02:50.650 They won't bother. Right. So the solution proposed 00:02:50.650 --> 00:02:55.009 is this idea of modular thinking and specialization. 00:02:55.330 --> 00:02:57.129 The analogy they use is pretty spot on. It's 00:02:57.129 --> 00:02:59.669 like building a human team. You don't hire one 00:02:59.669 --> 00:03:02.270 super person to do everything right. No, you 00:03:02.270 --> 00:03:05.030 hire specialists, a researcher, a writer, maybe 00:03:05.030 --> 00:03:08.319 a salesperson. Precisely. And the idea is that 00:03:08.319 --> 00:03:10.860 AI agents should be specialists too. Each one 00:03:10.860 --> 00:03:12.939 does one thing, but does it exceptionally well. 00:03:13.659 --> 00:03:15.560 And this gives you that dual benefit the source 00:03:15.560 --> 00:03:17.539 talks about. They become easy to build, easy 00:03:17.539 --> 00:03:20.280 to manage, and actually surprisingly reliable 00:03:20.280 --> 00:03:23.479 because they're so focused. So, okay, if we really 00:03:23.479 --> 00:03:25.919 strip away all those complex layers, what's the 00:03:25.919 --> 00:03:28.659 fundamental benefit? What's the core shift that 00:03:28.659 --> 00:03:32.159 this simplicity enables for us? The fundamental 00:03:32.159 --> 00:03:34.740 shift, really, is that simplicity makes these 00:03:34.740 --> 00:03:37.360 AI systems reliable, manageable, and maybe most 00:03:37.360 --> 00:03:40.300 importantly, accessible. Anyone can adopt them. 00:03:40.560 --> 00:03:42.979 It turns intimidating tech into a practical tool 00:03:42.979 --> 00:03:46.340 for daily use. It democratizes it. OK, so we 00:03:46.340 --> 00:03:49.219 see the power of keeping things simple, specializing. 00:03:49.740 --> 00:03:51.819 But what does that actually look like when you're 00:03:51.819 --> 00:03:54.379 designing one of these specific AI agents? Our 00:03:54.379 --> 00:03:56.340 source gives this great framework, calling it 00:03:56.340 --> 00:03:59.060 the anatomy of an effective agent. Yeah, that's 00:03:59.060 --> 00:04:00.840 a key point. What are the core components? Think 00:04:00.840 --> 00:04:04.159 of it just like writing up a super clear job 00:04:04.159 --> 00:04:06.199 description for a person. The source says these 00:04:06.199 --> 00:04:09.099 six elements pretty much determine 80 % of whether 00:04:09.099 --> 00:04:12.099 an agent will succeed or not. OK, first up, role 00:04:12.099 --> 00:04:14.520 and responsibility. This is about defining its 00:04:14.520 --> 00:04:17.879 exact task. And we mean specific, like market 00:04:17.879 --> 00:04:20.459 trend analysis specialist or email draft assistant. 00:04:20.860 --> 00:04:23.420 The narrower that role, the better it performs. 00:04:23.879 --> 00:04:26.920 It's all about singular focus. And then you got 00:04:26.920 --> 00:04:29.540 the tools. What outside resources can it actually 00:04:29.540 --> 00:04:32.040 use? This could be, say, perplexity for web search, 00:04:32.279 --> 00:04:34.480 maybe Google Workspace for handling documents, 00:04:34.699 --> 00:04:36.680 or Gmail for sending emails. These are like its 00:04:36.680 --> 00:04:39.100 digital hands and feet letting it interact. Next, 00:04:39.459 --> 00:04:42.060 input requirements. What information does the 00:04:42.060 --> 00:04:44.550 agent need just to get started? It could be simple, 00:04:44.689 --> 00:04:47.209 like a research topic or maybe a link to a Google 00:04:47.209 --> 00:04:49.709 Docs document. Clear expectations right from 00:04:49.709 --> 00:04:52.529 the start. Right. Following that, we have workflow 00:04:52.529 --> 00:04:54.730 details. These are the actual sequential steps 00:04:54.730 --> 00:04:57.470 the agent follows internally. For instance, one, 00:04:57.629 --> 00:05:00.829 get the topic. Two, run three searches. Three, 00:05:01.149 --> 00:05:03.810 synthesize those results. Four, create a report. 00:05:04.050 --> 00:05:07.160 It's that step -by -step process. Then boundaries. 00:05:07.519 --> 00:05:09.639 These are crucial. The rules it absolutely must 00:05:09.639 --> 00:05:12.360 not break. Things like do not give financial 00:05:12.360 --> 00:05:15.399 advice or maybe keep the report under a thousand 00:05:15.399 --> 00:05:18.980 words or always use a professional tone. These 00:05:18.980 --> 00:05:21.519 are the guardrails keeping it on track. Exactly, 00:05:21.839 --> 00:05:24.660 guardrails. And finally, output format. What 00:05:24.660 --> 00:05:26.639 should the end result actually look like? Is 00:05:26.639 --> 00:05:30.879 it a link to a new Google Doc? A JSON file? Or 00:05:30.879 --> 00:05:33.019 just a simple confirmation message saying an 00:05:33.019 --> 00:05:35.399 email has been sent. This gives you predictable 00:05:35.399 --> 00:05:39.279 results every single time. So why is being so 00:05:39.279 --> 00:05:41.000 precise about all these different components? 00:05:41.439 --> 00:05:43.680 Why is that so crucial for how the agent actually 00:05:43.680 --> 00:05:45.839 performs? What's the sort of hidden benefit of 00:05:45.839 --> 00:05:47.720 planning it out so meticulously? Well, defining 00:05:47.720 --> 00:05:50.060 these elements so clearly ensures the AI agent 00:05:50.060 --> 00:05:52.759 stays laser focused. It performs its single task 00:05:52.759 --> 00:05:54.920 exceptionally well, which drastically reduces 00:05:54.920 --> 00:05:57.180 errors and boosts reliability. Simple as that. 00:05:57.579 --> 00:05:59.860 OK, so we've seen why simplicity is key. what 00:05:59.860 --> 00:06:01.759 makes up one of these specialist agents. The 00:06:01.759 --> 00:06:04.279 next logical step is, how do you actually build 00:06:04.279 --> 00:06:06.379 them? And the source provides this surprisingly 00:06:06.379 --> 00:06:08.660 straightforward three -step framework, makes 00:06:08.660 --> 00:06:11.079 it very actionable. And it starts, interestingly, 00:06:11.279 --> 00:06:13.620 with observation. Before you even think about 00:06:13.620 --> 00:06:15.740 building, you first need to map out what you 00:06:15.740 --> 00:06:19.160 actually do day -to -day. Right. Step one, map 00:06:19.160 --> 00:06:22.449 your workflow. This means you just watch yourself 00:06:22.449 --> 00:06:25.730 work. Look for those repetitive, time -consuming 00:06:25.730 --> 00:06:29.089 tasks that follow a pretty clear process. The 00:06:29.089 --> 00:06:31.790 source suggests things like doing research, creating 00:06:31.790 --> 00:06:35.370 reports, maybe writing drafts or processing data. 00:06:36.050 --> 00:06:38.790 But the key here, it says, is to cleanly separate 00:06:38.790 --> 00:06:41.389 the tasks. Don't just lump research and write 00:06:41.389 --> 00:06:43.389 a blog post together. No, you split it. Split 00:06:43.389 --> 00:06:45.430 it. You'd have a research agent and then a separate 00:06:45.430 --> 00:06:47.889 writing agent. It's like taking a big overwhelming 00:06:47.889 --> 00:06:50.170 job and just breaking it down into smaller, more 00:06:50.170 --> 00:06:53.709 manageable pieces. Exactly. Then step two flows 00:06:53.709 --> 00:06:56.829 right from that. Design specialized AI assistance. 00:06:56.990 --> 00:06:59.389 Each of those cleanly separated tasks gets its 00:06:59.389 --> 00:07:02.290 own independent single purpose agent. So each 00:07:02.290 --> 00:07:04.269 agent is its own little workflow, totally focused 00:07:04.269 --> 00:07:06.490 on doing one thing really, really well. And then 00:07:06.490 --> 00:07:09.110 step three, add a manager agent. This is the 00:07:09.110 --> 00:07:11.689 orchestra conductor part. This manager, it doesn't 00:07:11.689 --> 00:07:14.250 actually do the specialized work itself. Instead, 00:07:14.290 --> 00:07:17.129 it takes your overall request. understands it, 00:07:17.430 --> 00:07:19.310 and then coordinates and delegates the workout 00:07:19.310 --> 00:07:22.170 to those specialist agents, just like a department 00:07:22.170 --> 00:07:24.689 head would manage a team of people. Yeah, perfect 00:07:24.689 --> 00:07:27.430 analogy. Now before you jump in, you do need 00:07:27.430 --> 00:07:30.160 to prep your tools a bit. you'll need an N8n 00:07:30.160 --> 00:07:33.019 account. The source describes this as a powerful 00:07:33.019 --> 00:07:36.980 no -code platform. Think of it like digital Legos 00:07:36.980 --> 00:07:39.680 for your tasks. It lets you visually connect 00:07:39.680 --> 00:07:42.420 apps and build these workflows. OK, digital Legos. 00:07:42.420 --> 00:07:45.620 I like that. Then you'll need API keys from AI 00:07:45.620 --> 00:07:48.540 services. These are like unique digital fingerprints, 00:07:48.699 --> 00:07:50.800 right? They give you access to specific AI models. 00:07:50.959 --> 00:07:54.269 Yep. Like OpenAI for GPT -4 or maybe Anthropic 00:07:54.269 --> 00:07:56.910 for their Claude 3 models, the source highlights 00:07:56.910 --> 00:07:58.949 Sonnet as a good balance between performance 00:07:58.949 --> 00:08:01.670 and cost. Got it. You'll also need a search service 00:08:01.670 --> 00:08:03.889 perplexity API key, as mentioned, so your agents 00:08:03.889 --> 00:08:05.410 can actually access the internet for current 00:08:05.410 --> 00:08:07.970 info. And finally, Google Workspace access for 00:08:07.970 --> 00:08:10.290 docs and Gmail, so they can read, write, and 00:08:10.290 --> 00:08:13.310 send things. And you'll set up specific credentials 00:08:13.310 --> 00:08:15.509 for secure access, but the platform guides you 00:08:15.509 --> 00:08:20.709 through that. So this whole framework... It fundamentally 00:08:20.709 --> 00:08:23.569 simplifies what often feels like incredibly complex 00:08:23.569 --> 00:08:26.050 automation, doesn't it? What's the biggest barrier 00:08:26.050 --> 00:08:28.189 it kind of breaks down for the average person 00:08:28.189 --> 00:08:30.189 wanting to try this? Yeah, it breaks down that 00:08:30.189 --> 00:08:32.429 intimidating complexity into these manageable 00:08:32.429 --> 00:08:35.090 specialized parts. It makes sophisticated AI 00:08:35.090 --> 00:08:38.190 accessible for clear execution, even if you can't 00:08:38.190 --> 00:08:40.830 code. OK, let's get practical now. This is where 00:08:40.830 --> 00:08:43.070 it gets really interesting. Our source walks 00:08:43.070 --> 00:08:46.230 us through building your very first AI assistant. 00:08:46.590 --> 00:08:49.409 a market research specialist. Super common task, 00:08:49.730 --> 00:08:52.009 right? Research a topic, summarize it in a report. 00:08:52.169 --> 00:08:55.450 Right. And the soul of any agent, as the source 00:08:55.450 --> 00:08:57.750 calls it, is its system prompt. But here's the 00:08:57.750 --> 00:09:00.970 brilliant part, the hack. The source suggests 00:09:00.970 --> 00:09:03.529 using another AI to actually design your prompt 00:09:03.529 --> 00:09:06.009 for you. That's actually pretty meta. I like 00:09:06.009 --> 00:09:08.850 it. Very clever. Yeah. So for step one... Design 00:09:08.850 --> 00:09:10.669 the system prompt you'd literally go to something 00:09:10.669 --> 00:09:13.570 like chat GPT or Claude and you'd ask it to draft 00:09:13.570 --> 00:09:15.570 a detailed system prompt You'd say my agents 00:09:15.570 --> 00:09:18.690 role is market research specialist. It needs 00:09:18.690 --> 00:09:21.909 to use perplexity for search Access this Google 00:09:21.909 --> 00:09:25.090 Doc for company context create a detailed analysis 00:09:25.090 --> 00:09:28.029 report in a new Google Doc and give me the link 00:09:28.490 --> 00:09:31.169 You even tell it the structure you want. Executive 00:09:31.169 --> 00:09:33.730 summary, key analysis, opportunities, challenges. 00:09:33.929 --> 00:09:36.309 And the AI helper just generates that detailed 00:09:36.309 --> 00:09:38.230 prompt. It's like having an expert assistant 00:09:38.230 --> 00:09:40.250 for your assistant. Saves a ton of guesswork. 00:09:40.470 --> 00:09:44.779 Okay, then step two. Create the workflow in NAN. 00:09:45.200 --> 00:09:47.539 You name it something clear, like Agent One Research 00:09:47.539 --> 00:09:49.960 Specialist. You add what's called an AI agent 00:09:49.960 --> 00:09:52.279 node. Think of this as the dedicated brain with 00:09:52.279 --> 00:09:54.820 an ANN just for this one agent. Right, it's where 00:09:54.820 --> 00:09:57.100 the core intelligence sits. You configure the 00:09:57.100 --> 00:10:00.100 AI model there, maybe GPT 4 .1, and you add memory. 00:10:00.399 --> 00:10:02.740 The source suggests simple memory, maybe five 00:10:02.740 --> 00:10:05.100 messages context. This is like giving the AI 00:10:05.100 --> 00:10:07.340 a very short -term working memory. Just enough 00:10:07.340 --> 00:10:09.139 to remember the last few bits of the conversation 00:10:09.139 --> 00:10:11.500 so it stays on track without getting, you know, 00:10:11.639 --> 00:10:14.639 confused. In step three, add the necessary tools. 00:10:14.820 --> 00:10:17.340 You add the perplexity search tool, connect your 00:10:17.340 --> 00:10:20.139 API key, and actually let the AI figure out the 00:10:20.139 --> 00:10:22.820 best search queries itself. You add a Google 00:10:22.820 --> 00:10:25.860 Docs read tool, link your Google account, Point 00:10:25.860 --> 00:10:27.899 it to your company info doc for that context. 00:10:28.080 --> 00:10:30.480 And then a Google Docs Create tool. You set permissions 00:10:30.480 --> 00:10:33.120 and even let the AI design the document titles. 00:10:33.360 --> 00:10:36.139 For step four, add the system prompt. You just 00:10:36.139 --> 00:10:38.539 copy paste that prompt your AI helper generated 00:10:38.539 --> 00:10:41.720 right into the AI agent node. Then you can fine 00:10:41.720 --> 00:10:43.820 tune it a bit, maybe specify the report lengths, 00:10:44.379 --> 00:10:47.080 like around 800 words, or really emphasize that 00:10:47.080 --> 00:10:49.200 section structure again. Make it even sharper. 00:10:49.480 --> 00:10:51.740 And then step five, test your research agent. 00:10:51.909 --> 00:10:54.629 You activate it, give it a real request, something 00:10:54.629 --> 00:10:57.049 like, please research AI in the retail sector 00:10:57.049 --> 00:10:59.809 in Vietnam, focusing on practical applications, 00:11:00.190 --> 00:11:02.070 pioneering companies, and forecasts for the next 00:11:02.070 --> 00:11:04.169 three years, and relate this to our company's 00:11:04.169 --> 00:11:06.070 product development strategy. And then you watch. 00:11:06.509 --> 00:11:08.590 The agent should sequentially perform all those 00:11:08.590 --> 00:11:11.669 steps. Search, read the company doc, create the 00:11:11.669 --> 00:11:14.009 report, and finally give you back the link. And 00:11:14.009 --> 00:11:16.440 if something goes wrong? Debugging is simple. 00:11:16.899 --> 00:11:18.919 NET has this executions panel that shows you 00:11:18.919 --> 00:11:21.240 exactly what happened at each step and where 00:11:21.240 --> 00:11:23.620 it failed if it did. Makes fixing things much 00:11:23.620 --> 00:11:26.000 easier. It really is like giving an assistant 00:11:26.000 --> 00:11:29.259 a very specific brief and all the tool they need 00:11:29.259 --> 00:11:31.820 to get the job done, isn't it? What's the immediate 00:11:31.820 --> 00:11:33.899 payoff you see from setting it up with that level 00:11:33.899 --> 00:11:36.519 of clarity? Exactly. You define the role, give 00:11:36.519 --> 00:11:39.539 the tools, set clear rules. The immediate payoff 00:11:39.539 --> 00:11:42.200 is highly predictable, consistent results. You 00:11:42.200 --> 00:11:43.799 know what you're going to get. Okay. Agent one 00:11:43.799 --> 00:11:46.590 is built. The source then moves on to agent two, 00:11:46.950 --> 00:11:49.110 the data presentation designer. This one takes 00:11:49.110 --> 00:11:52.409 that potentially dense research report from agent 00:11:52.409 --> 00:11:55.169 one and turns it into something much more visual, 00:11:55.210 --> 00:11:58.110 more consumable, like a blog outline with image 00:11:58.110 --> 00:12:00.929 suggestions and then emails it out. Yeah, this 00:12:00.929 --> 00:12:03.049 agent really showcases the versatility here. 00:12:03.149 --> 00:12:05.370 It's not just about generating text from scratch. 00:12:05.450 --> 00:12:07.350 It's about transforming information from one 00:12:07.350 --> 00:12:10.169 format into a completely new, maybe more valuable 00:12:10.169 --> 00:12:13.330 one. So for step one. Design the system prompt 00:12:13.330 --> 00:12:15.210 for this agent. You go back to your AI prompt 00:12:15.210 --> 00:12:17.330 designer again. Your little helper AI, yeah. 00:12:17.610 --> 00:12:19.409 You'd ask it to write a prompt for a content 00:12:19.409 --> 00:12:22.870 designer agent. Tell it. This agent gets a Google 00:12:22.870 --> 00:12:25.490 Docs link as input. It needs to read the doc, 00:12:25.929 --> 00:12:28.070 create a detailed blog post outline based on 00:12:28.070 --> 00:12:31.169 it, suggest visual aids like bar charts, infographics, 00:12:31.429 --> 00:12:33.610 stock photos for each section, and then compose 00:12:33.610 --> 00:12:35.809 and send a professional email with that outline 00:12:35.809 --> 00:12:39.220 to a specific address. Nice and clear. Then, 00:12:39.360 --> 00:12:42.779 step two, set up the workflow in N8M. You create 00:12:42.779 --> 00:12:45.419 Agent 2 Content Designer. Again, you add that 00:12:45.419 --> 00:12:48.399 AI agent node, probably using GPT 4 .1 again. 00:12:48.799 --> 00:12:52.039 But here's a key tip from the source. For creative 00:12:52.039 --> 00:12:54.720 tasks that involve a lot of text, like drafting 00:12:54.720 --> 00:12:57.600 outlines or emails, you might want to increase 00:12:57.600 --> 00:13:00.820 the token limit, maybe up to around 9 ,000 tokens. 00:13:00.960 --> 00:13:02.480 OK, tokens. Can you break that down quickly? 00:13:02.580 --> 00:13:04.960 What does increasing the limit actually do? Sure. 00:13:05.120 --> 00:13:08.419 Think of tokens as like words or parts of words 00:13:08.419 --> 00:13:11.120 that the AI processes. Increasing the limit basically 00:13:11.120 --> 00:13:13.340 gives the AI more working space or short -term 00:13:13.340 --> 00:13:15.840 memory. It lets it handle longer inputs, think 00:13:15.840 --> 00:13:18.039 more deeply, and generate longer, more complex 00:13:18.039 --> 00:13:19.860 outputs without forgetting the beginning of its 00:13:19.860 --> 00:13:21.799 thought. Got it. More mental bandwidth. Exactly. 00:13:22.139 --> 00:13:23.899 And you'd add a simple memory tool here too. 00:13:24.259 --> 00:13:27.340 Okay. Then in step three, configure the tools, 00:13:27.720 --> 00:13:29.840 you add the Google Docs Read tool again. But 00:13:29.840 --> 00:13:32.620 this time, the AI just uses the link you provide 00:13:32.620 --> 00:13:35.360 to figure out which document to read. Then you 00:13:35.360 --> 00:13:37.539 add a Gmail send tool, configure the permissions, 00:13:37.980 --> 00:13:40.019 and let the AI design the subject line and the 00:13:40.019 --> 00:13:42.460 whole email body itself. You're trusting it more. 00:13:43.100 --> 00:13:46.059 And for step four, add system instructions, you 00:13:46.059 --> 00:13:48.399 paste in your generated prompt, and again, customize 00:13:48.399 --> 00:13:51.240 it. Maybe specify the blog's tone, friendly yet 00:13:51.240 --> 00:13:53.960 professional, or clarify the email format, use 00:13:53.960 --> 00:13:55.720 bullet points for the outline, for instance. 00:13:56.139 --> 00:13:59.440 Makes sense. And finally, step five, test the 00:13:59.440 --> 00:14:02.100 visual agent. You run the workflow, give it a 00:14:02.100 --> 00:14:04.360 request like, please create a blog post outline 00:14:04.360 --> 00:14:07.080 from this research report, paste the Google Docs 00:14:07.080 --> 00:14:09.539 link you got from Agent 1, and email the result 00:14:09.539 --> 00:14:12.519 to me at example at email .com. And off it goes. 00:14:12.759 --> 00:14:15.059 Reads the doc, creates the outline with visual 00:14:15.059 --> 00:14:18.360 ideas, composes the email, and sends it. Job 00:14:18.360 --> 00:14:20.679 done. So this agent isn't just spitting out text, 00:14:20.740 --> 00:14:22.879 it's structuring information, it's thinking visually, 00:14:23.200 --> 00:14:25.379 it's even managing communication by sending the 00:14:25.379 --> 00:14:27.919 email. How does this kind of transformation capability 00:14:27.919 --> 00:14:30.129 really change the game for productivity? Oh, 00:14:30.190 --> 00:14:33.070 it changes everything. It transforms that raw 00:14:33.070 --> 00:14:35.090 data into something much more digestible, more 00:14:35.090 --> 00:14:37.509 visually oriented. It streamlines the entire 00:14:37.509 --> 00:14:39.850 content creation process from initial research 00:14:39.850 --> 00:14:42.389 right through to distribution planning. Midroll 00:14:42.389 --> 00:14:45.289 sponsor, Read. Right, this next part. This is 00:14:45.289 --> 00:14:47.700 where the real magic seems to happen. Our source 00:14:47.700 --> 00:14:50.059 calls it the orchestra conductor. It's where 00:14:50.059 --> 00:14:52.600 these independent specialized agents stop being 00:14:52.600 --> 00:14:55.639 just individual tools and become a true collaborative 00:14:55.639 --> 00:14:57.820 team. Yeah, if we kind of zoom out and connect 00:14:57.820 --> 00:15:00.019 this to the bigger picture, this is about creating 00:15:00.019 --> 00:15:02.759 a genuine workflow, not just automating isolated 00:15:02.759 --> 00:15:05.740 tasks. Well, it's automation on a whole different 00:15:05.740 --> 00:15:09.200 level, really. So step one, prepare your specialist 00:15:09.200 --> 00:15:11.799 agents. Before you even build the manager, you 00:15:11.799 --> 00:15:13.740 need to tweak those first two agents slightly. 00:15:13.929 --> 00:15:16.889 It's a small change, but vital. You go into their 00:15:16.889 --> 00:15:18.889 settings in N and N and change their trigger 00:15:18.889 --> 00:15:22.070 node from on chat to when execute by another 00:15:22.070 --> 00:15:25.470 workflow. Right. And doing that generates a special 00:15:25.470 --> 00:15:28.289 webhook URL for each agent. Think of it like 00:15:28.289 --> 00:15:30.649 a unique digital doorbell that only the manager 00:15:30.649 --> 00:15:33.149 agent will know how to ring. You also add an 00:15:33.149 --> 00:15:35.570 input field so they can receive specific commands 00:15:35.570 --> 00:15:38.230 from the manager. OK, doorbell analogy works. 00:15:38.730 --> 00:15:40.450 And then there's a crucial point about memory. 00:15:40.700 --> 00:15:43.200 Yes, absolutely crucial. You remove the local 00:15:43.200 --> 00:15:45.879 memory node from the specialist agents, agent 00:15:45.879 --> 00:15:48.740 one and agent two. Why? Why take away their memory? 00:15:49.039 --> 00:15:50.879 Because the manager agent is going to handle 00:15:50.879 --> 00:15:53.220 all the conversation memory for the entire team 00:15:53.220 --> 00:15:55.879 interaction. This stops the context from getting 00:15:55.879 --> 00:15:57.860 fragmented. It ensures the whole process feels 00:15:57.860 --> 00:16:00.879 like one single coherent conversation managed 00:16:00.879 --> 00:16:03.419 from the top down. Keeps everything seamless. 00:16:03.580 --> 00:16:07.419 Got it. Centralized memory. Then step two, create 00:16:07.419 --> 00:16:09.889 the manager workflow. You make a new workflow, 00:16:10.269 --> 00:16:12.830 we call it manager -agent -ai -team -lead. You 00:16:12.830 --> 00:16:15.629 add an AI agent node again. Here, the source 00:16:15.629 --> 00:16:17.950 suggests you might want a more powerful model, 00:16:18.269 --> 00:16:20.750 like Claude III Opus, if the coordination is 00:16:20.750 --> 00:16:23.429 complex, though Sonnet is often still very capable. 00:16:23.769 --> 00:16:25.450 And crucially, you add that simple memory node 00:16:25.450 --> 00:16:27.289 here on the manager. This is vital because it 00:16:27.289 --> 00:16:29.330 needs to track your initial request, which agent 00:16:29.330 --> 00:16:31.409 it called, what the result was, what the next 00:16:31.409 --> 00:16:33.690 step is, the whole coordination process. Makes 00:16:33.690 --> 00:16:38.000 sense. Next, step three, connect your team. Inside 00:16:38.000 --> 00:16:40.740 the manager agent's workflow, you add the call 00:16:40.740 --> 00:16:44.240 and aid workflow tool. It's basically a webhook 00:16:44.240 --> 00:16:46.299 node configured to make requests. You'll add 00:16:46.299 --> 00:16:48.460 two of these. One for your research agent. You 00:16:48.460 --> 00:16:51.659 paste in its unique webhook URL, its doorbell 00:16:51.659 --> 00:16:54.240 address, and give it a clear name inside the 00:16:54.240 --> 00:16:56.580 manager's tool settings, like research specialist 00:16:56.580 --> 00:16:59.000 tool. And another one for your content agent, 00:16:59.440 --> 00:17:01.399 pasting its unique URL and naming it something 00:17:01.399 --> 00:17:03.799 like... content designer tool. Yeah, that clean 00:17:03.799 --> 00:17:05.720 naming is really important because it helps the 00:17:05.720 --> 00:17:08.640 manager AI understand exactly which tool corresponds 00:17:08.640 --> 00:17:10.920 to which specialist agent's function when you 00:17:10.920 --> 00:17:13.460 define its instructions. Okay, which brings us 00:17:13.460 --> 00:17:15.980 to step four, design the manager system prompt. 00:17:16.190 --> 00:17:17.970 This is probably the most important prompt of 00:17:17.970 --> 00:17:20.450 all because it defines how the manager thinks 00:17:20.450 --> 00:17:22.970 and acts as the coordinator. It's the brain of 00:17:22.970 --> 00:17:24.609 the operation. You'll tell it something like, 00:17:25.190 --> 00:17:28.049 you are an effective AI team lead. Your job is 00:17:28.049 --> 00:17:30.349 to coordinate a team of specialist agents. You 00:17:30.349 --> 00:17:32.710 have access to the following tools. Research 00:17:32.710 --> 00:17:35.789 specialist tool and content designer tool. Then 00:17:35.789 --> 00:17:38.490 you give it the logic. Right. When the user gives 00:17:38.490 --> 00:17:41.509 a request, analyze it carefully. If the request 00:17:41.509 --> 00:17:43.869 requires both research and content creation, 00:17:44.170 --> 00:17:46.740 you must follow these steps precisely. First, 00:17:47.059 --> 00:17:49.400 call the Research Specialist tool with the research 00:17:49.400 --> 00:17:51.680 task. Wait for its response, which will be a 00:17:51.680 --> 00:17:54.059 document link. Then, call the Content Designer 00:17:54.059 --> 00:17:56.180 tool, providing that document link. Finally, 00:17:56.559 --> 00:17:59.019 report back to the user when the entire process 00:17:59.019 --> 00:18:00.960 is complete. You're programming its decision 00:18:00.960 --> 00:18:03.849 -making process. And finally, step five. Test 00:18:03.849 --> 00:18:07.009 your complete AI team. You open the chat interface 00:18:07.009 --> 00:18:09.269 for the manager agent itself, and you give it 00:18:09.269 --> 00:18:11.490 that composite request, the multi -step one, 00:18:11.690 --> 00:18:14.450 like, please research the future of remote work 00:18:14.450 --> 00:18:16.650 and then create a blog outline from the research 00:18:16.650 --> 00:18:19.869 results and email it to me. Whoa. Just imagine 00:18:19.869 --> 00:18:22.470 watching this unfold in the NN interface. The 00:18:22.470 --> 00:18:24.390 manager gets the request. It figures out it needs 00:18:24.390 --> 00:18:26.809 agent one first, calls the research specialist 00:18:26.809 --> 00:18:29.309 tool. Agent one runs, does its thing, creates 00:18:29.309 --> 00:18:31.730 the report, sends the link back. Manager receives 00:18:31.730 --> 00:18:33.910 the link, then it knows, OK, next step, Agent 00:18:33.910 --> 00:18:36.490 2. Calls the content designer tool, passing that 00:18:36.490 --> 00:18:38.609 link along. Agent 2 runs, reads the doc, creates 00:18:38.609 --> 00:18:40.890 the outline, sends the email. And then the manager 00:18:40.890 --> 00:18:43.289 agent just calmly reports back to you, OK, I've 00:18:43.289 --> 00:18:45.430 completed the research and emailed the blog outline 00:18:45.430 --> 00:18:48.349 as requested, it feels like. Like you've just 00:18:48.349 --> 00:18:50.609 built a tiny automated work supply chain right 00:18:50.609 --> 00:18:54.329 there. And imagining scaling this, maybe not 00:18:54.329 --> 00:18:57.690 a billion queries, but scaling it up. The possibilities 00:18:57.690 --> 00:18:59.609 feel kind of limitless, don't they? They really 00:18:59.609 --> 00:19:02.109 do. It's a powerful concept. So this manager 00:19:02.109 --> 00:19:05.589 agent truly acts as the brain, the central intelligence 00:19:05.589 --> 00:19:08.150 directing the whole operation. It turns those 00:19:08.150 --> 00:19:10.809 isolated tools into a properly coordinated effort. 00:19:11.089 --> 00:19:13.650 Yes, precisely. It orchestrates the specialized 00:19:13.650 --> 00:19:16.130 agents, ensuring a seamless automated workflow 00:19:16.130 --> 00:19:18.809 where individual tasks combine into a unified, 00:19:19.029 --> 00:19:21.680 efficient process. OK, so we've built our basic 00:19:21.680 --> 00:19:24.099 team. Now, how do we level them up, make them 00:19:24.099 --> 00:19:26.819 even better, more robust? And how do we start 00:19:26.819 --> 00:19:29.799 applying this framework more broadly across the 00:19:29.799 --> 00:19:32.059 different kinds of tasks? What are the next steps 00:19:32.059 --> 00:19:34.319 for really optimizing these systems? Yeah, that's 00:19:34.319 --> 00:19:36.400 a really important question. How do we ensure 00:19:36.400 --> 00:19:38.940 they're reliable enough for real world use? And 00:19:38.940 --> 00:19:41.220 how do we adapt them? It boils down to refinement 00:19:41.220 --> 00:19:43.740 and building and resilience. OK, first point, 00:19:44.180 --> 00:19:47.230 making agents more reliable. The source really 00:19:47.230 --> 00:19:50.509 hammers on using specific prawns. The more detailed 00:19:50.509 --> 00:19:53.049 your instructions, the less likely you are to 00:19:53.049 --> 00:19:56.289 get those weird AI hallucinations or off -track 00:19:56.289 --> 00:19:59.529 results. And it mentions few -shot prompting 00:19:59.529 --> 00:20:02.609 as a way to improve quality. Now, I'll admit... 00:20:02.359 --> 00:20:04.700 I still wrestle with prompt drift myself sometimes, 00:20:04.980 --> 00:20:08.140 where the AI subtly changes its output style 00:20:08.140 --> 00:20:11.079 over time. Specificity here really does feel 00:20:11.079 --> 00:20:13.480 key. But for people new to it, could you quickly 00:20:13.480 --> 00:20:15.359 explain Fuchsha? What kind of examples would 00:20:15.359 --> 00:20:17.410 you give it? Absolutely. Few -shot prompting 00:20:17.410 --> 00:20:19.789 is essentially like showing the AI a couple of 00:20:19.789 --> 00:20:21.529 really good examples of what you want before 00:20:21.529 --> 00:20:23.910 you give it the actual task. So instead of just 00:20:23.910 --> 00:20:25.970 describing the output, you provide maybe two 00:20:25.970 --> 00:20:29.190 or three pairs of sample inputs and the corresponding 00:20:29.190 --> 00:20:31.369 perfect outputs you're looking for. Ah, OK, like 00:20:31.369 --> 00:20:33.470 showing it to completed homework. Exactly. It 00:20:33.470 --> 00:20:35.630 shows the AI the precise pattern, the style, 00:20:35.769 --> 00:20:38.269 the level of detail you expect. This dramatically 00:20:38.269 --> 00:20:40.349 reduces the chance of those hallucinations or 00:20:40.349 --> 00:20:42.470 the kind of prompt drift you mentioned, because 00:20:42.470 --> 00:20:45.509 it has concrete examples to mimic. Show, don't 00:20:45.509 --> 00:20:48.210 just tell. That makes so much sense. OK. And 00:20:48.210 --> 00:20:51.410 the source also stresses. Test in isolation. 00:20:52.170 --> 00:20:54.250 Always make sure your specialist agents work 00:20:54.250 --> 00:20:56.150 perfectly on their own before you connect them 00:20:56.150 --> 00:20:58.950 to the manager. Much easier to debug that way. 00:20:59.329 --> 00:21:01.970 Right. And error handling. Adding instructions 00:21:01.970 --> 00:21:04.390 to the manager's prompt. Yeah. Simple things 00:21:04.390 --> 00:21:07.329 like, if a tool fails when you call it, try calling 00:21:07.329 --> 00:21:09.390 it one more time. If it still fails, don't just 00:21:09.390 --> 00:21:11.569 stop. Report the specific error back to the user. 00:21:12.150 --> 00:21:14.519 That builds in a crucial layer of resilience, 00:21:14.980 --> 00:21:17.460 makes it less brittle. And what about scaling 00:21:17.460 --> 00:21:21.299 your AI team? How easy is it to add more specialists? 00:21:21.609 --> 00:21:23.930 That's the beauty of this modular design. It's 00:21:23.930 --> 00:21:26.130 built for scalability. Let's say you want to 00:21:26.130 --> 00:21:28.470 add Agent 3 social media specialist. You just 00:21:28.470 --> 00:21:31.190 build its dedicated workflow, get its webhook 00:21:31.190 --> 00:21:34.289 URL. Add a new call N8n workflow tool in the 00:21:34.289 --> 00:21:36.410 manager. Update the manager system prompt so 00:21:36.410 --> 00:21:38.390 it knows about its new team member and when to 00:21:38.390 --> 00:21:40.589 call it. And then test the whole integrated team. 00:21:40.910 --> 00:21:43.309 Exactly. It's remarkably straightforward to expand 00:21:43.309 --> 00:21:45.390 the team's capabilities. Then there's introducing 00:21:45.390 --> 00:21:48.029 the human in the loop. This is important, right? 00:21:48.289 --> 00:21:52.029 Not everything should be or can be 100 % automated. 00:21:52.650 --> 00:21:54.990 Absolutely not. And it's easy to build in checks. 00:21:55.069 --> 00:21:57.569 For example, after the research agent creates 00:21:57.569 --> 00:21:59.789 the report, the manager could send you the link 00:21:59.789 --> 00:22:02.349 via chat and ask, the research report is ready. 00:22:02.769 --> 00:22:04.670 Would you like me to proceed with creating the 00:22:04.670 --> 00:22:08.180 blog outline now? And it waits. It waits. The 00:22:08.180 --> 00:22:11.099 process only continues if you reply yes or give 00:22:11.099 --> 00:22:13.500 some other confirmation. You can do this with 00:22:13.500 --> 00:22:15.900 things called wait nodes in NAN or just building 00:22:15.900 --> 00:22:18.000 conditional logic into the manager's prompt. 00:22:18.240 --> 00:22:21.420 It becomes collaboration, not just line automation. 00:22:22.039 --> 00:22:24.400 You stay in control. That's really powerful. 00:22:24.480 --> 00:22:26.839 And the real -world applications mentioned, they 00:22:26.839 --> 00:22:29.509 really drive home the potential. Marketing, trend 00:22:29.509 --> 00:22:32.309 research plus content writing plus social media 00:22:32.309 --> 00:22:34.390 scheduling agents working together. Or sales, 00:22:34.890 --> 00:22:37.609 lead prospecting plus proposal drafting plus 00:22:37.609 --> 00:22:39.990 follow -up email agents. Imagine that sequence 00:22:39.990 --> 00:22:43.210 automated. Customer service, an FAQ agent handling 00:22:43.210 --> 00:22:45.890 common queries, a triage agent routing complex 00:22:45.890 --> 00:22:48.289 issues, and an escalation agent flagging things 00:22:48.289 --> 00:22:50.519 for human support. The possibilities really start 00:22:50.519 --> 00:22:53.059 to open up, freeing up so much human time and 00:22:53.059 --> 00:22:55.519 energy for higher level tasks. So boiling it 00:22:55.519 --> 00:22:57.200 all down, what's the single biggest takeaway 00:22:57.200 --> 00:22:59.859 for someone listening right now wanting to actually 00:22:59.859 --> 00:23:02.319 leverage this framework to make a real impact 00:23:02.319 --> 00:23:05.000 on their daily work? Start small. Don't try to 00:23:05.000 --> 00:23:07.980 boil the ocean. Pick one or two specialized tasks, 00:23:08.220 --> 00:23:10.559 build those agents, and then coordinate them 00:23:10.559 --> 00:23:12.880 with a simple manager agent. That's the path 00:23:12.880 --> 00:23:15.759 to scalable, effective automation that actually 00:23:15.759 --> 00:23:18.319 works. OK. So reflecting on all this, what does 00:23:18.319 --> 00:23:20.900 it really mean for you, the listener? The core 00:23:20.900 --> 00:23:23.880 idea from this deep dive seems crystal clear. 00:23:24.759 --> 00:23:27.980 building effective AI teams. It doesn't have 00:23:27.980 --> 00:23:30.339 to be this overwhelming, super technical mountain 00:23:30.339 --> 00:23:32.500 to climb. Not at all. The golden rules we've 00:23:32.500 --> 00:23:35.400 seen are simplicity, specialization, and intelligent 00:23:35.400 --> 00:23:38.299 coordination. You break down those complex jobs 00:23:38.299 --> 00:23:41.720 into single, manageable pieces. Give each AI 00:23:41.720 --> 00:23:43.859 agent just one focused task, let it get really 00:23:43.859 --> 00:23:46.039 good at that, and then have a manager agent lead 00:23:46.039 --> 00:23:48.740 them, orchestrating the whole workflow. And this 00:23:48.740 --> 00:23:51.339 isn't about replacing human ingenuity or connection, 00:23:51.380 --> 00:23:53.359 is it? It feels more like equipping ourselves 00:23:53.359 --> 00:23:55.670 with incredible powerful, reliable assistants. 00:23:55.990 --> 00:23:58.609 That's exactly it. They handle the routine, the 00:23:58.609 --> 00:24:00.809 repetitive stuff, the tedious bits that drain 00:24:00.809 --> 00:24:04.710 our energy, and that frees you up to focus on 00:24:04.710 --> 00:24:06.849 the strategic thinking, the real creativity, 00:24:07.009 --> 00:24:09.130 the problem -solving, and building those essential 00:24:09.130 --> 00:24:11.690 human connections. So are you ready to build 00:24:11.690 --> 00:24:14.210 your own AI team? We'd really encourage you to 00:24:14.210 --> 00:24:16.700 start with just one agent. Maybe that research 00:24:16.700 --> 00:24:18.980 specialist we walk through. Yeah, just try building 00:24:18.980 --> 00:24:21.859 one. You'll likely be genuinely amazed at the 00:24:21.859 --> 00:24:23.880 transformation even that single step can bring 00:24:23.880 --> 00:24:26.619 to your workflow and your focus. We definitely 00:24:26.619 --> 00:24:28.940 encourage you to explore the source material. 00:24:29.160 --> 00:24:31.640 If you want more of the nitty -gritty, the detailed 00:24:31.640 --> 00:24:34.519 steps, the future of work, it really is here. 00:24:34.740 --> 00:24:36.819 And perhaps surprisingly, it can be built on 00:24:36.819 --> 00:24:39.299 a foundation of simplicity. Thanks so much for 00:24:39.299 --> 00:24:41.519 joining us for this deep dive into building AI 00:24:41.519 --> 00:24:44.500 agent teams. Until next time, keep learning and 00:24:44.500 --> 00:24:45.180 keep exploring.