WEBVTT 00:00:00.000 --> 00:00:01.120 You know, you build these things, right? You 00:00:01.120 --> 00:00:04.540 spend time putting together an AI agent in N8N, 00:00:04.719 --> 00:00:06.599 hoping it's going to, you know, automate something 00:00:06.599 --> 00:00:10.480 cool for you. Yeah. And then it just goes completely 00:00:10.480 --> 00:00:13.800 off the rails. It like makes stuff up or maybe 00:00:13.800 --> 00:00:16.000 it totally ignores the tools you gave it and 00:00:16.000 --> 00:00:20.359 your whole workflow just breaks. Yes. That feeling. 00:00:20.420 --> 00:00:22.359 It's incredibly frustrating. Oh, absolutely. 00:00:22.660 --> 00:00:25.399 Anyone who's tried to move AI from a cool demo 00:00:25.399 --> 00:00:30.000 to actual automation has hit that wall. And I 00:00:30.000 --> 00:00:33.560 think a big part of it is just how a lot of resources 00:00:33.560 --> 00:00:35.759 still teach you to build these things. Right. 00:00:35.799 --> 00:00:38.219 It feels like the standard amateur way you see 00:00:38.219 --> 00:00:41.359 online is, what, grab a basic model, write a 00:00:41.359 --> 00:00:44.259 super simple prompt, cross your fingers, and 00:00:44.259 --> 00:00:46.520 just hope for the best. Pretty much. And then 00:00:46.520 --> 00:00:48.719 you're surprised when it's not reliable. Exactly. 00:00:48.719 --> 00:00:50.280 Exactly. And the source material we've got for 00:00:50.280 --> 00:00:52.039 this deep dive, it really cuts through all that. 00:00:52.100 --> 00:00:53.939 It's pretty blunt about it, saying if that's 00:00:53.939 --> 00:00:55.759 what you're doing, you're kind of building a 00:00:55.759 --> 00:01:00.820 toy, not a production system. Ooh. Okay. So this 00:01:00.820 --> 00:01:03.520 deep dive then, this is like the antidote. It's 00:01:03.520 --> 00:01:07.760 about moving past the toy phase. actually building 00:01:07.760 --> 00:01:09.519 something that's robust, something you can depend 00:01:09.519 --> 00:01:11.680 on. That's the mission. It's about pulling back 00:01:11.680 --> 00:01:14.299 the curtain on what the professionals really 00:01:14.299 --> 00:01:17.000 focus on. It's less about just prompt engineering, 00:01:17.260 --> 00:01:18.920 although that's still important, and way more 00:01:18.920 --> 00:01:21.400 about the configuration. The configuration. Yeah. 00:01:21.420 --> 00:01:24.140 The source highlights seven key settings that, 00:01:24.180 --> 00:01:27.219 honestly, most people just seem to miss or maybe 00:01:27.219 --> 00:01:30.780 ignore. Seven. Wow. Okay. So we're talking like... 00:01:31.420 --> 00:01:33.579 architectural design here, right? Building predictable, 00:01:33.799 --> 00:01:36.659 reliable systems, not just, you know, getting 00:01:36.659 --> 00:01:39.340 lucky with some clever wording in a prompt. Precisely. 00:01:39.340 --> 00:01:42.219 This deep dive is all about understanding the 00:01:42.219 --> 00:01:44.980 engine control unit for your AI automation, not 00:01:44.980 --> 00:01:47.319 just the driver trying to steer. Okay. I love 00:01:47.319 --> 00:01:49.480 that. The engine control unit. So, all right, 00:01:49.480 --> 00:01:52.120 let's unpack this. Where do we even start with 00:01:52.120 --> 00:01:54.719 this configuration mindset? Like, why does it 00:01:54.719 --> 00:01:57.159 matter so much? Well, you know, the source makes 00:01:57.159 --> 00:01:59.459 it really clear from the jump. Configuration 00:01:59.459 --> 00:02:01.780 is the foundation. It's the difference between 00:02:01.780 --> 00:02:04.120 something that's maybe a cheap internal tool 00:02:04.120 --> 00:02:07.680 that saves you an hour here and there and a mission 00:02:07.680 --> 00:02:09.460 -critical business system that you actually depend 00:02:09.460 --> 00:02:12.560 on. Missing these settings means, as they put 00:02:12.560 --> 00:02:14.340 it, your brilliant prompt is like a brilliant 00:02:14.340 --> 00:02:16.860 driver in a car that's just fundamentally not 00:02:16.860 --> 00:02:18.560 configured right. It's never going to perform 00:02:18.560 --> 00:02:21.840 reliably. Okay, so it's not like a minor tweak. 00:02:22.099 --> 00:02:24.319 It's the core of building something that's not 00:02:24.319 --> 00:02:26.599 just going to... I don't know, randomly decided 00:02:26.599 --> 00:02:29.340 to potato halfway through your process. Right. 00:02:29.400 --> 00:02:32.740 It's about predictability and scale. And the 00:02:32.740 --> 00:02:35.360 source dives into these settings one by one, 00:02:35.379 --> 00:02:38.080 which is super helpful. OK, let's do it. First 00:02:38.080 --> 00:02:40.979 up, they talk about choosing the right AI brain. 00:02:41.060 --> 00:02:44.159 Yeah. And I got to confess, defaulting to like 00:02:44.159 --> 00:02:47.419 the GPT -4 or the latest famous model for everything 00:02:47.419 --> 00:02:49.419 that feels relatable. I think I've done that. 00:02:49.520 --> 00:02:51.080 Yeah. And the source uses this great analogy. 00:02:51.560 --> 00:02:54.879 Using GPT -4 for every single task is like using 00:02:54.879 --> 00:02:57.639 a Formula One race car to deliver a pizza. I 00:02:57.639 --> 00:02:59.759 mean, it's incredibly powerful, sure, but it's 00:02:59.759 --> 00:03:03.120 also wildly expensive and frankly, totally overkill 00:03:03.120 --> 00:03:06.080 for, you know, a simple delivery job. It's the 00:03:06.080 --> 00:03:08.400 wrong tool. OK, yeah, that makes so much sense. 00:03:08.539 --> 00:03:10.960 So what's the what's the professional framework 00:03:10.960 --> 00:03:14.560 for picking the right brain for the right job? 00:03:14.800 --> 00:03:16.840 The source lays it out really clearly based on 00:03:16.840 --> 00:03:19.219 the task you're trying to accomplish. If you 00:03:19.219 --> 00:03:21.400 need what they call Einstein level reasoning, 00:03:21.520 --> 00:03:23.780 complex problem solving, strategic thinking, 00:03:23.979 --> 00:03:26.520 deep synthesis, you're looking at models like 00:03:26.520 --> 00:03:32.080 maybe DeepSeek V2 or some of OpenAI's newer O1 00:03:32.080 --> 00:03:35.340 series models. A good use case there would be 00:03:35.340 --> 00:03:37.960 like analyzing complex market trends to develop 00:03:37.960 --> 00:03:41.060 a strategic plan. But heads up, these really 00:03:41.060 --> 00:03:42.960 heavy thinkers, some of them might not support 00:03:42.960 --> 00:03:45.259 tools well. Check the docs on that. Okay. So 00:03:45.259 --> 00:03:47.599 for the really big, chewy problems, what about 00:03:47.599 --> 00:03:51.379 speed, like real -time interaction? Ah, speed 00:03:51.379 --> 00:03:53.319 is totally different. For light and fast responses, 00:03:53.680 --> 00:03:55.979 real -time chat, interactive apps, you need models 00:03:55.979 --> 00:03:58.379 optimized for pure speed. Yeah. Think Grok or 00:03:58.379 --> 00:04:00.639 maybe Gemini 2 .5 Flats. Gemini Flats, yeah. 00:04:00.719 --> 00:04:02.879 That's your customer support chatbot on a website. 00:04:03.120 --> 00:04:05.199 It needs to feel instantaneous, right? Got it. 00:04:05.580 --> 00:04:08.439 Speed demons for speed tasks. And what about 00:04:08.439 --> 00:04:11.340 bigger companies with, you know, strict security 00:04:11.340 --> 00:04:14.240 needs, compliance and all that? That's where 00:04:14.240 --> 00:04:16.839 you lean into enterprise -grade security and 00:04:16.839 --> 00:04:19.379 governance, models available via services like 00:04:19.379 --> 00:04:22.500 Azure OpenAI or AWS Bedrock. Right, the cloud 00:04:22.500 --> 00:04:25.180 provider wrappers. Exactly. If you're dealing 00:04:25.180 --> 00:04:29.339 with, say, IPA or GDPR compliance, or you just 00:04:29.339 --> 00:04:31.319 need to integrate deeply into a secure cloud 00:04:31.319 --> 00:04:34.079 environment, using these services provides the 00:04:34.079 --> 00:04:38.310 power of... models like GPT -4, but within that 00:04:38.310 --> 00:04:41.129 necessary secure wrapper. Like for HR data or 00:04:41.129 --> 00:04:43.449 something? Yeah. An internal HR assistant dealing 00:04:43.449 --> 00:04:45.930 with sensitive employee data is a classic example. 00:04:46.230 --> 00:04:49.279 Okay. What if privacy is paramount? Like the 00:04:49.279 --> 00:04:51.839 data absolutely cannot leave my servers. Then 00:04:51.839 --> 00:04:53.779 you're going privacy first and self -hosting. 00:04:53.899 --> 00:04:56.300 Models run locally using something like Ollama, 00:04:56.420 --> 00:04:59.399 maybe running Lama3 or Mistral locally. Or using 00:04:59.399 --> 00:05:01.680 Mistral models directly gives you absolute data 00:05:01.680 --> 00:05:03.800 control, which is huge for some businesses. Right. 00:05:03.819 --> 00:05:05.660 If you're dealing with super sensitive IP or 00:05:05.660 --> 00:05:08.660 financials. Exactly. Processing, you know, highly 00:05:08.660 --> 00:05:10.980 proprietary financial data on a local server, 00:05:11.120 --> 00:05:14.160 for instance. What about seeing images or like 00:05:14.160 --> 00:05:16.800 reading charts, multimodal stuff? Yep. For that 00:05:16.800 --> 00:05:19.089 multimodal magic, you need... specific models 00:05:19.089 --> 00:05:22.850 Google Gemini 2 .5 Pro is known for strong vision 00:05:22.850 --> 00:05:26.410 capabilities or maybe GPT 4 .5 fear in that ecosystem 00:05:26.410 --> 00:05:29.149 okay analyzing images of product damage from 00:05:29.149 --> 00:05:32.129 support tickets seeing the photo and understanding 00:05:32.129 --> 00:05:34.490 the text description that's a perfect use case 00:05:34.490 --> 00:05:36.930 there okay so we've got power speed security 00:05:36.930 --> 00:05:41.209 privacy vision that's a whole lineup but what 00:05:41.209 --> 00:05:44.980 about cost because Honestly, those fancy models 00:05:44.980 --> 00:05:47.259 can get really expensive, right? Massively. And 00:05:47.259 --> 00:05:49.279 that's where the cost optimized production models 00:05:49.279 --> 00:05:52.480 become essential. Think GPT -4 mini or other 00:05:52.480 --> 00:05:54.959 lightweight models. When the task is relatively 00:05:54.959 --> 00:05:57.839 simple, like basic classification, simple data 00:05:57.839 --> 00:06:00.560 extraction or Q &A, and you need to process thousands 00:06:00.560 --> 00:06:03.279 or millions of requests affordably. These models 00:06:03.279 --> 00:06:06.620 are often, you know. 90 % as capable as their 00:06:06.620 --> 00:06:08.899 bigger siblings, but at a tiny fraction of the 00:06:08.899 --> 00:06:11.220 cost. Okay, wow. So the real professional approach 00:06:11.220 --> 00:06:13.500 isn't just picking one favorite model, but like 00:06:13.500 --> 00:06:15.639 having a whole arsenal and choosing the absolute 00:06:15.639 --> 00:06:18.550 best one for each query on the fly. Exactly. 00:06:18.689 --> 00:06:20.709 And that's the dynamic model selection they talk 00:06:20.709 --> 00:06:23.230 about, often using a model router. A model router. 00:06:23.430 --> 00:06:26.689 Yeah. Instead of hard coding, say, GPT -4 for 00:06:26.689 --> 00:06:29.629 every single thing, you use a much cheaper, much 00:06:29.629 --> 00:06:32.629 faster AI agent, that's your router, whose only 00:06:32.629 --> 00:06:35.410 job is to read the user's query or the workflow's 00:06:35.410 --> 00:06:38.170 need and decide where of those other more specialized 00:06:38.170 --> 00:06:40.649 models is the most appropriate and cost effective. 00:06:40.930 --> 00:06:44.170 Wait, so you use one AI to pick which AI to use. 00:06:44.209 --> 00:06:47.439 That's kind of meta. It is, but it's incredibly 00:06:47.439 --> 00:06:50.279 efficient at scale. The router agent uses a simple 00:06:50.279 --> 00:06:52.860 system prompt with decision rules based on the 00:06:52.860 --> 00:06:55.620 query type. Is it complex reasoning? Is it asking 00:06:55.620 --> 00:06:58.000 for code? Is it general chat? Is it something 00:06:58.000 --> 00:07:00.899 simple? Is it multimodal? Then the real worker 00:07:00.899 --> 00:07:03.959 agent uses an expression in AN to dynamically 00:07:03.959 --> 00:07:06.319 call the specific model the router recommended. 00:07:06.720 --> 00:07:08.779 You might even use a service like OpenRouter 00:07:08.779 --> 00:07:10.360 that connects to a whole bunch of different models. 00:07:10.459 --> 00:07:12.439 Okay, whoa. So you actually get the best model 00:07:12.439 --> 00:07:15.279 for the specific job and you keep costs way down. 00:07:15.399 --> 00:07:18.060 because you're not using that Ferrari to deliver 00:07:18.060 --> 00:07:20.660 every single email. That really makes so much 00:07:20.660 --> 00:07:22.860 sense for a production system. It's a hallmark 00:07:22.860 --> 00:07:25.439 of a robust production system, yeah. Finding 00:07:25.439 --> 00:07:27.579 that perfect balance of capability and cost. 00:07:27.860 --> 00:07:30.240 All right, so picking the right brain or dynamically 00:07:30.240 --> 00:07:33.480 routing to the right brain is clearly key. What's 00:07:33.480 --> 00:07:36.579 next on the list of settings people often overlook? 00:07:37.069 --> 00:07:40.089 Controlling creativity and reliability. This 00:07:40.089 --> 00:07:43.389 is where temperature and top P come in. And honestly, 00:07:43.629 --> 00:07:45.569 leaving these at their defaults is just like 00:07:45.569 --> 00:07:48.290 asking for unpredictable results. Temperature. 00:07:48.730 --> 00:07:50.670 I've seen that slider in interfaces. What does 00:07:50.670 --> 00:07:53.589 it actually do? Think of temperature as the creativity 00:07:53.589 --> 00:07:56.230 and risk dial, usually from 0 .0 up to like 1 00:07:56.230 --> 00:07:59.910 .0 or more. A low temperature, like 0 .1 to 0 00:07:59.910 --> 00:08:02.389 .3, means the AI is going to stick to the most 00:08:02.389 --> 00:08:05.410 common, safest, most probable word choices. Okay. 00:08:05.509 --> 00:08:08.269 It's super predictable, very factual, very consistent. 00:08:08.889 --> 00:08:11.610 That's what you want for, say, a support chatbot 00:08:11.610 --> 00:08:13.990 that needs to give the exact same correct answer 00:08:13.990 --> 00:08:16.129 every single time someone asks about your return 00:08:16.129 --> 00:08:18.379 policy. Okay. So predictable and boring. Boring, 00:08:18.459 --> 00:08:21.160 almost. What if you want something more engaging, 00:08:21.459 --> 00:08:24.980 more creative? Then you go medium, maybe 0 .6 00:08:24.980 --> 00:08:28.939 to 0 .9. The AI takes more risks, uses less common 00:08:28.939 --> 00:08:30.839 words, feels much more creative and human -like. 00:08:30.939 --> 00:08:32.820 That's perfect for marketing copy, generating 00:08:32.820 --> 00:08:35.940 social media posts, brainstorming ideas. And 00:08:35.940 --> 00:08:38.940 high. Yeah. Like 1 .0 and above. What happens 00:08:38.940 --> 00:08:41.759 then? That's the danger zone. It becomes extremely 00:08:41.759 --> 00:08:45.059 unpredictable, often random, sometimes nonsensical 00:08:45.059 --> 00:08:48.240 or, you know, wildly hallucinatory. It's almost 00:08:48.240 --> 00:08:50.360 never used in production unless you're like trying 00:08:50.360 --> 00:08:53.159 to brainstorm abstract art concepts with zero 00:08:53.159 --> 00:08:55.940 need for accuracy whatsoever. Got it. So temperature 00:08:55.940 --> 00:08:58.419 is kind of like how adventurous the AI is with 00:08:58.419 --> 00:09:00.779 its word choices. What about top P? How does 00:09:00.779 --> 00:09:02.580 that work? Does it relate? It does. Yeah. Top 00:09:02.580 --> 00:09:05.080 P works with temperature. Think of it as the 00:09:05.080 --> 00:09:07.940 word choice filter or the size of the pool of 00:09:07.940 --> 00:09:10.450 words the AI. considers for the next word. A 00:09:10.450 --> 00:09:13.669 low top P, say 0 .3, means it only looks at the 00:09:13.669 --> 00:09:17.330 top 30 % most likely words. Very safe, very conservative 00:09:17.330 --> 00:09:20.409 choices. And a high top P, like 0 .9. A much 00:09:20.409 --> 00:09:22.789 wider range of words. Like 0 .9 means it considers 00:09:22.789 --> 00:09:25.889 the top 90 % most likely words. That allows for 00:09:25.889 --> 00:09:28.809 a lot more variety and creativity. A top P of 00:09:28.809 --> 00:09:32.690 1 .0 considers all possible words, which can 00:09:32.690 --> 00:09:35.190 lead to some pretty bizarre outputs. So low temp 00:09:35.190 --> 00:09:37.289 and low top P is super conservative. High temp 00:09:37.289 --> 00:09:40.980 and high top. p is pretty much chaos pretty much 00:09:40.980 --> 00:09:43.259 the source gives this really handy production 00:09:43.259 --> 00:09:45.720 settings formula table as a starting point for 00:09:45.720 --> 00:09:48.200 different use cases for reliable business automation 00:09:48.200 --> 00:09:51.779 they suggest temperature around 0 .2 top p maybe 00:09:51.779 --> 00:09:55.580 0 .7 okay low temp medium -ish top p yeah and 00:09:55.580 --> 00:09:58.019 often adding a frequency penalty of 0 .5 to 1 00:09:58.019 --> 00:10:00.799 .0 to prevent the ai from repeating itself that's 00:10:00.799 --> 00:10:02.879 low creativity high reliability low repetition 00:10:02.879 --> 00:10:05.330 and for like generating creative content Marketing 00:10:05.330 --> 00:10:08.289 stuff. 10 .7, top P 1 .0, frequency penalty maybe 00:10:08.289 --> 00:10:11.009 0 .5, high creativity, wide word choice, keep 00:10:11.009 --> 00:10:13.730 it fresh. And like super structured output like 00:10:13.730 --> 00:10:16.330 JSON or code where format is critical. You need 00:10:16.330 --> 00:10:19.029 extremely low creativity there to ensure it follows 00:10:19.029 --> 00:10:22.309 the format precisely. 10 .1, top P lower 0 .5, 00:10:22.450 --> 00:10:24.309 maybe frequency penalty 0 .2. You're forcing 00:10:24.309 --> 00:10:26.809 it into a narrow, precise path. Okay, so you 00:10:26.809 --> 00:10:29.169 really, you have to dial these in for each specific 00:10:29.169 --> 00:10:31.250 task. You can't just leave them at default and 00:10:31.250 --> 00:10:34.200 hope it works. Crucial. And the source also quickly 00:10:34.200 --> 00:10:36.259 mentions other critical model settings you find 00:10:36.259 --> 00:10:38.940 in that node. Max tokens, which prevents rambling 00:10:38.940 --> 00:10:41.580 and controls costs. Right, stop it talking forever. 00:10:41.899 --> 00:10:44.519 Timeout, for how long the workflow waits for 00:10:44.519 --> 00:10:47.700 a response. And max retries, which is important 00:10:47.700 --> 00:10:50.200 for production to handle transient API failures. 00:10:50.379 --> 00:10:54.000 Set that to like two or three. Good tip. Okay, 00:10:54.279 --> 00:10:56.200 so we've covered picking the brain and controlling 00:10:56.200 --> 00:10:59.100 its personality, I guess, with temperature and 00:10:59.100 --> 00:11:02.240 top P. What about memory? AI agents need to remember 00:11:02.240 --> 00:11:04.399 past turns in a conversation, right? Yes. And 00:11:04.399 --> 00:11:06.299 this is another place where the amateur move 00:11:06.299 --> 00:11:08.919 can absolutely kill you in production. Using 00:11:08.919 --> 00:11:11.879 the standard NANN simple memory or window buffer 00:11:11.879 --> 00:11:14.759 memory nodes. Why? They seem easy to use when 00:11:14.759 --> 00:11:17.460 you're testing. They are easy for simple tests 00:11:17.460 --> 00:11:20.740 or internal only tools with maybe one user, but 00:11:20.740 --> 00:11:23.320 they use your server's RAM in production with, 00:11:23.419 --> 00:11:26.320 you know. potentially dozens or hundreds of users 00:11:26.320 --> 00:11:28.840 all having separate conversations, that memory 00:11:28.840 --> 00:11:31.700 usage just grows and grows and grows. Eventually, 00:11:31.879 --> 00:11:34.600 it'll exhaust your server's RAM, slowing everything 00:11:34.600 --> 00:11:37.399 down, and ultimately crashing your entire NN 00:11:37.399 --> 00:11:40.399 instance. Oh, okay. It's fundamentally not scalable 00:11:40.399 --> 00:11:42.820 for multi -user scenarios. Okay, so that's like 00:11:42.820 --> 00:11:45.480 a hidden scaling killer. What's the professional 00:11:45.480 --> 00:11:48.759 alternative then? Using a dedicated, external, 00:11:48.980 --> 00:11:52.080 scalable database for chat memory. The source 00:11:52.080 --> 00:11:55.240 specifically recommends Postgresql, and they 00:11:55.240 --> 00:11:58.000 point to Supabase as a really good option because 00:11:58.000 --> 00:12:00.200 it's easy to get started with and has a free 00:12:00.200 --> 00:12:03.440 tier. Supabase, okay. I've heard of them. How 00:12:03.440 --> 00:12:06.440 does that setup work? Is it complicated? According 00:12:06.440 --> 00:12:08.340 to the source, it's pretty straightforward. You 00:12:08.340 --> 00:12:10.419 create a free Supabase account, and then in your 00:12:10.419 --> 00:12:12.200 project settings, you get a connection string. 00:12:12.360 --> 00:12:14.179 They mention the transaction pillar connection 00:12:14.179 --> 00:12:15.840 string is important for managing connections 00:12:15.840 --> 00:12:18.000 efficiently. Cooler. Got it. You just copy those 00:12:18.000 --> 00:12:21.970 details, host name, database name, user. password 00:12:21.970 --> 00:12:25.409 into the NANN Postgres chat memory credential 00:12:25.409 --> 00:12:28.389 setup. So NANN talks directly to that database 00:12:28.389 --> 00:12:31.169 to store and retrieve the chat history. Exactly. 00:12:31.529 --> 00:12:35.409 And crucially, in the NETEN AI agent node itself, 00:12:35.470 --> 00:12:38.590 when you configure the memory, you must set a 00:12:38.590 --> 00:12:42.509 unique session ID for each user. Ah, so conversations 00:12:42.509 --> 00:12:45.350 don't get mixed up. Precisely. This is how the 00:12:45.350 --> 00:12:47.549 database knows which conversation history belongs 00:12:47.549 --> 00:12:50.889 to whom. You can use their username, email, or 00:12:50.889 --> 00:12:53.950 a unique ID from your system. And you give the 00:12:53.950 --> 00:12:56.629 table in the database a name, like NAT histories 00:12:56.629 --> 00:12:58.889 or whatever makes sense. Okay, that makes total 00:12:58.889 --> 00:13:01.450 sense. Separate histories, stored scalably in 00:13:01.450 --> 00:13:04.110 a database, not blowing up my server RAM. And 00:13:04.110 --> 00:13:06.250 you also configure the context window length 00:13:06.250 --> 00:13:08.289 here, right? How many past messages it remembers. 00:13:08.509 --> 00:13:10.750 Yes, that's configured in the node as well. That 00:13:10.750 --> 00:13:13.370 controls how many of the most recent messages 00:13:13.370 --> 00:13:15.950 the agent remembers and sends back to the AI 00:13:15.950 --> 00:13:18.250 model with each turn in the conversation. Right. 00:13:18.330 --> 00:13:20.029 How do you decide how long that should be? For 00:13:20.029 --> 00:13:22.450 short, simple chats like basic customer support 00:13:22.450 --> 00:13:25.429 FAQs, maybe five, 10 messages is enough. And 00:13:25.429 --> 00:13:28.029 for longer, more complex stuff like planning 00:13:28.029 --> 00:13:30.750 or research. Yeah. For conversations that build 00:13:30.750 --> 00:13:32.970 over time, like project planning or analytical 00:13:32.970 --> 00:13:36.210 tasks, you might need 20 or more messages for 00:13:36.210 --> 00:13:38.850 the AI to maintain good continuity and context. 00:13:39.269 --> 00:13:42.029 So the rule of thumb is a larger window means 00:13:42.029 --> 00:13:44.870 better memory, but also higher cost and potentially 00:13:44.870 --> 00:13:47.970 slower response. Precisely. Sending more text 00:13:47.970 --> 00:13:50.370 with every single request means a higher API 00:13:50.370 --> 00:13:52.929 cost and more processing time for the model. 00:13:53.049 --> 00:13:55.570 You absolutely have to tune this setting based 00:13:55.570 --> 00:13:58.029 on what your specific use case actually needs. 00:13:58.110 --> 00:14:01.090 Don't just crank it up to max. Okay. Memory problem 00:14:01.090 --> 00:14:04.940 solved with the database. Next up, tools. Giving 00:14:04.940 --> 00:14:07.259 agents the ability to use tools seems pretty 00:14:07.259 --> 00:14:09.399 fundamental. What's the configuration secret 00:14:09.399 --> 00:14:11.799 there? The golden rule, as the source puts it. 00:14:11.879 --> 00:14:14.320 The more fields you leave blank for the AI to 00:14:14.320 --> 00:14:16.779 dynamically define when it calls a tool, the 00:14:16.779 --> 00:14:19.019 higher the chance of hallucination, malformed 00:14:19.019 --> 00:14:22.000 data, and the tool call failing. Oof, that resonates. 00:14:22.320 --> 00:14:23.860 Can you give an example? Like, how does that 00:14:23.860 --> 00:14:25.879 play out? Sure. They use the classic send email 00:14:25.879 --> 00:14:28.620 tool the wrong way. You leave the to email address, 00:14:28.820 --> 00:14:31.179 the subject, and the message body entirely up 00:14:31.179 --> 00:14:32.840 to the AI to figure out from the conversation. 00:14:33.200 --> 00:14:35.919 Right. Just tell it, send an email about X to 00:14:35.919 --> 00:14:38.600 Bob. And the AI might invent an email address 00:14:38.600 --> 00:14:40.799 for Bob that doesn't exist, forget the subject 00:14:40.799 --> 00:14:43.019 line entirely, or mix up the recipient and the 00:14:43.019 --> 00:14:45.759 content. Disaster. Yeah, I've definitely seen 00:14:45.759 --> 00:14:49.200 variations of that happen. The right way. Predefine 00:14:49.200 --> 00:14:51.379 everything you possibly can in the tool node 00:14:51.379 --> 00:14:54.240 itself. Only let the AI control the specific 00:14:54.240 --> 00:14:57.240 parts where its creativity or understanding of 00:14:57.240 --> 00:14:59.840 the conversation is actually needed. Oh, okay. 00:14:59.960 --> 00:15:02.159 So for the email tool, you'd use an expression 00:15:02.159 --> 00:15:05.279 to pull the to email address from, say, a previous 00:15:05.279 --> 00:15:08.200 workflow step or a database lookup. You'd only 00:15:08.200 --> 00:15:10.820 let the AI define the subject. and the message 00:15:10.820 --> 00:15:13.700 dynamically based on the user's request. Okay, 00:15:13.779 --> 00:15:16.440 so you're essentially constraining the AI's creativity 00:15:16.440 --> 00:15:18.779 to only the parts of the tool call where its 00:15:18.779 --> 00:15:21.139 linguistic intelligence is valuable, not where 00:15:21.139 --> 00:15:23.580 it can just invent data and break the tool. Exactly. 00:15:23.639 --> 00:15:26.159 It dramatically increases the reliability of 00:15:26.159 --> 00:15:28.519 your tool calls, and the source lists the available 00:15:28.519 --> 00:15:32.259 tool types in AN8n, the hundreds of pre -built 00:15:32.259 --> 00:15:34.620 nodes for common services like Gmail, Slack, 00:15:34.919 --> 00:15:36.960 Airtable, Notion, databases. That would be usual 00:15:36.960 --> 00:15:40.860 suspects. Yeah. Then the universal HTTP request 00:15:40.860 --> 00:15:44.860 node for interacting with literally any API and 00:15:44.860 --> 00:15:47.980 even custom tools, which let you wrap another 00:15:47.980 --> 00:15:50.659 any and end workflow as a tool the AI can call, 00:15:50.799 --> 00:15:53.080 which is pretty powerful. Pro tip is probably 00:15:53.080 --> 00:15:55.399 always start with pre -built. Absolutely. They're 00:15:55.399 --> 00:15:58.159 the most robust and easiest to configure. Only 00:15:58.159 --> 00:16:01.559 resort to custom or HTTP if there's truly no 00:16:01.559 --> 00:16:03.860 pre -built node that does what you need. All 00:16:03.860 --> 00:16:05.899 right. We've picked the right brain. We've controlled 00:16:05.899 --> 00:16:08.159 its randomness. We've given it scalable memory 00:16:08.159 --> 00:16:11.659 and configured its tools smartly by predefining 00:16:11.659 --> 00:16:14.100 fields. What about the prompt itself? You said 00:16:14.100 --> 00:16:15.860 it's not just prompting, but it's still a key 00:16:15.860 --> 00:16:17.899 piece, right? It is, but it's about writing a 00:16:17.899 --> 00:16:20.500 professional system prompt, not just, you know, 00:16:20.500 --> 00:16:22.679 be a helpful assistant. Like that vague stuff. 00:16:22.820 --> 00:16:24.399 Yeah, that's like hiring someone and just saying, 00:16:24.460 --> 00:16:26.700 hey, be helpful without any other direction. 00:16:27.279 --> 00:16:29.159 The source provides a really solid structured 00:16:29.159 --> 00:16:31.580 template, which is key. Okay, walk me through 00:16:31.580 --> 00:16:33.500 that professional prompt template. What are the 00:16:33.500 --> 00:16:35.379 essential sections? It breaks it down logically. 00:16:35.539 --> 00:16:37.720 First, you have the role persona. What is this 00:16:37.720 --> 00:16:40.659 AI, EG? You are a helpful customer service assistant 00:16:40.659 --> 00:16:43.980 for ACME Corp. Clear identity. Then the primary 00:16:43.980 --> 00:16:47.580 goal. What is its main task? EGG, your main task 00:16:47.580 --> 00:16:50.379 is to provide prompt, accurate, and friendly 00:16:50.379 --> 00:16:53.240 responses to customer inquiries based on the 00:16:53.240 --> 00:16:55.799 provided tools and knowledge. Okay, clear job 00:16:55.799 --> 00:16:59.019 title and objective. Right. Then, domain knowledge 00:16:59.019 --> 00:17:01.960 context. This is where you inject specific facts 00:17:01.960 --> 00:17:04.339 about your business, products, policies, etc. 00:17:04.740 --> 00:17:07.180 This is crucial so it doesn't have to guess or 00:17:07.180 --> 00:17:09.440 hallucinate information about your company. Ah, 00:17:09.680 --> 00:17:11.480 okay, so it actually knows what it's talking 00:17:11.480 --> 00:17:13.940 about regarding my business. That's huge. Exactly. 00:17:14.220 --> 00:17:17.640 Then, tools. Explicitly list the tools it has 00:17:17.640 --> 00:17:20.119 access to, when and why it should use them, and 00:17:20.119 --> 00:17:23.400 maybe a rule like, prioritize using your tools 00:17:23.400 --> 00:17:25.880 over guessing if information might exist outside 00:17:25.880 --> 00:17:28.420 your core knowledge. Tell it what tools it has 00:17:28.420 --> 00:17:31.119 and how to use them. Yeah. And if you have a 00:17:31.119 --> 00:17:33.380 knowledge -based search tool, like for searching 00:17:33.380 --> 00:17:35.880 documentation, clearly lay out the rules for 00:17:35.880 --> 00:17:39.089 using it. Use this tool for all factual questions, 00:17:39.349 --> 00:17:41.849 base your responses strictly on the search results, 00:17:42.029 --> 00:17:44.970 and if the search tool finds no relevant information, 00:17:45.450 --> 00:17:47.769 state clearly that you could not find the answer 00:17:47.769 --> 00:17:50.269 in your knowledge base. Don't let it guess after 00:17:50.269 --> 00:17:52.769 a failed search. Okay, wow, so you're explicitly 00:17:52.769 --> 00:17:55.230 telling it what it knows, what it can do, and 00:17:55.230 --> 00:17:57.589 how it should use those capabilities. Very specific 00:17:57.589 --> 00:18:01.750 instructions. Yes. Then, format rules. What format 00:18:01.750 --> 00:18:05.289 should the final output be in? Plain text, markdown, 00:18:05.549 --> 00:18:08.329 JSON, maybe even a maximum word count. Tell it 00:18:08.329 --> 00:18:10.450 how to structure the answer. Style tone. How 00:18:10.450 --> 00:18:13.250 should it sound? AG, maintain a friendly and 00:18:13.250 --> 00:18:15.890 concise tone. Use the user's name if available. 00:18:16.089 --> 00:18:18.170 And safety and accuracy. That feels really important. 00:18:18.509 --> 00:18:21.069 Super critical section. You explicitly state 00:18:21.069 --> 00:18:23.430 rules like, if you are uncertain about an answer, 00:18:23.509 --> 00:18:25.049 state your uncertainty rather than guessing. 00:18:25.309 --> 00:18:27.809 Never disclose internal tool names or proprietary 00:18:27.809 --> 00:18:31.019 internal information. Always adhere to company 00:18:31.019 --> 00:18:33.680 policies and legal requirements. It's like giving 00:18:33.680 --> 00:18:36.220 the AI its full onboarding packet and employee 00:18:36.220 --> 00:18:40.119 handbook. All the guardrails. Pretty much. Finally, 00:18:40.220 --> 00:18:43.059 an optional but often useful one, reasoning. 00:18:43.579 --> 00:18:46.400 You can instruct it to think step -by -step internally 00:18:46.400 --> 00:18:49.759 before providing your final response. This encourages 00:18:49.759 --> 00:18:52.140 a more deliberate process, though you don't necessarily 00:18:52.140 --> 00:18:54.500 show these internal steps to the user. Okay, 00:18:54.539 --> 00:18:56.039 yeah, that makes so much more sense than just 00:18:56.039 --> 00:18:58.720 a couple of vague sentences. It moves from a... 00:18:58.970 --> 00:19:02.269 a vague command to a real structured professional 00:19:02.269 --> 00:19:05.130 delegation. It just dramatically improves reliability 00:19:05.130 --> 00:19:08.349 because the AI understands its boundaries, its 00:19:08.349 --> 00:19:11.309 resources, and the expected behavior. Less guessing, 00:19:11.450 --> 00:19:13.410 more structure. We've covered a lot of ground. 00:19:13.769 --> 00:19:15.950 Choosing the brain, controlling its output style, 00:19:16.190 --> 00:19:18.809 scalable memory, smart tools, professional prompts. 00:19:19.170 --> 00:19:21.549 What's next after the AI node itself? There's 00:19:21.549 --> 00:19:23.950 more. Output parsers. This solves that incredibly 00:19:23.950 --> 00:19:26.029 common problem where the AI gives you the correct 00:19:26.029 --> 00:19:27.970 answer, but it puts it in the wrong format for 00:19:27.970 --> 00:19:30.109 the next step in your innate end workflow. Oh, 00:19:30.109 --> 00:19:32.980 yeah. You need JSON. And it gives you a bulleted 00:19:32.980 --> 00:19:35.680 list. Or it messes up the JSON structure slightly. 00:19:35.859 --> 00:19:38.019 So frustrating. Happens all the time. Exactly. 00:19:38.259 --> 00:19:40.539 Output parsers are like the strict formatting 00:19:40.539 --> 00:19:43.500 guards right after the AI agent node in your 00:19:43.500 --> 00:19:46.460 workflow. The source lists three main types that 00:19:46.460 --> 00:19:48.740 are really useful. The structured output parser 00:19:48.740 --> 00:19:50.900 is for when you need a fixed JSON structure. 00:19:51.180 --> 00:19:54.039 You define the schema you expect. And this parser 00:19:54.039 --> 00:19:56.640 forces the AI's output into that exact shape, 00:19:56.720 --> 00:19:59.079 failing if it can't. Okay, so that guarantees 00:19:59.079 --> 00:20:02.500 my JSON format, mostly. What about lists of things? 00:20:02.660 --> 00:20:06.079 Like if I ask for five ideas. Item list output 00:20:06.079 --> 00:20:08.920 parser. This is for when the AI is supposed to 00:20:08.920 --> 00:20:11.200 return a variable list of items. You specify 00:20:11.200 --> 00:20:12.819 the separator character. Maybe it's AI between 00:20:12.819 --> 00:20:15.559 items or just a new line. And this parser takes 00:20:15.559 --> 00:20:18.440 that list -like text output and formats it into 00:20:18.440 --> 00:20:22.819 a clean usable array in N8n. Nice. Okay. And 00:20:22.819 --> 00:20:24.460 the third type, the one they call the professional 00:20:24.460 --> 00:20:27.329 safety net. What's that about? That's the autofixing 00:20:27.329 --> 00:20:29.049 output parser. This is for mission -critical 00:20:29.049 --> 00:20:31.210 workflows where a formatting error is just unacceptable. 00:20:31.430 --> 00:20:34.730 If your primary AI gives you broken JSON or some 00:20:34.730 --> 00:20:36.809 other malformed output... Which happens. Right. 00:20:36.930 --> 00:20:40.009 This parser makes a second... Usually very lightweight 00:20:40.009 --> 00:20:43.069 and cheap AI call. It sends the broken output 00:20:43.069 --> 00:20:45.369 to the secondary model with a simple instruction 00:20:45.369 --> 00:20:48.730 like, please fix this broken JSON. And it returns 00:20:48.730 --> 00:20:52.029 the corrected valid output. Wow. So if the main 00:20:52.029 --> 00:20:54.670 brain messes up the format, a backup brain just 00:20:54.670 --> 00:20:57.289 swoops in and fixes just that specific formatting 00:20:57.289 --> 00:21:00.480 problem. Exactly. It adds a tiny bit of cost 00:21:00.480 --> 00:21:03.200 and latency because it's a second API call, but 00:21:03.200 --> 00:21:05.319 it provides incredible resilience for systems 00:21:05.319 --> 00:21:07.799 that just cannot afford a formatting failure 00:21:07.799 --> 00:21:10.460 downstream. That's smart. The production strategy, 00:21:10.640 --> 00:21:12.660 they suggest, is often to use the structured 00:21:12.660 --> 00:21:15.799 parser for fixed formats and for critical workflows. 00:21:16.099 --> 00:21:18.779 Wrap that inside an auto -fixing parser node. 00:21:19.160 --> 00:21:21.519 Double protection. That is genuinely smart. It's 00:21:21.519 --> 00:21:24.039 like a little formatting quality control step 00:21:24.039 --> 00:21:26.259 that happens automatically. Built -in error correction. 00:21:26.480 --> 00:21:28.400 Yeah, it's essential for ensuring the AI's output 00:21:28.400 --> 00:21:30.839 integrates seamlessly into the rest of your automated 00:21:30.839 --> 00:21:33.039 processes. Stops the whole workflow breaking 00:21:33.039 --> 00:21:35.359 because of one misplaced bracket. All right, 00:21:35.400 --> 00:21:37.359 I think we've covered six of the seven. What's 00:21:37.359 --> 00:21:39.000 the last category of settings? Where else do 00:21:39.000 --> 00:21:41.619 we need to look? The agent level execution settings 00:21:41.619 --> 00:21:45.619 right there within the NANN AI agent node itself. 00:21:45.940 --> 00:21:48.519 The big one here is max iteration. Max iterations. 00:21:48.900 --> 00:21:50.960 What does that control exactly? This limits the 00:21:50.960 --> 00:21:53.619 number of back and forth steps the agent takes. 00:21:54.000 --> 00:21:56.940 You know, the cycle of the model thinking, deciding 00:21:56.940 --> 00:21:59.539 to use a tool, calling the tool, the tool returning 00:21:59.539 --> 00:22:01.720 a result, the model reading the result, thinking 00:22:01.720 --> 00:22:03.940 again, maybe calling another tool and so on. 00:22:04.339 --> 00:22:07.160 Model tool model. Okay, the thinking loop. Right. 00:22:07.240 --> 00:22:10.230 The default is often... 10 iterations. So if 00:22:10.230 --> 00:22:12.210 you have a pretty simple chat where it just needs 00:22:12.210 --> 00:22:14.589 to think and respond, 10 steps is probably fine. 00:22:14.730 --> 00:22:17.849 Probably. But for a complex analysis test that 00:22:17.849 --> 00:22:19.809 might involve multiple tool calls or breaking 00:22:19.809 --> 00:22:22.269 down a problem, you might need to increase that 00:22:22.269 --> 00:22:25.990 to 15 or 20. An autonomous research agent using 00:22:25.990 --> 00:22:28.470 several tools could potentially go up to 30. 00:22:28.650 --> 00:22:30.769 And the warning sign here, what's the danger? 00:22:31.230 --> 00:22:33.849 Setting max iterations too high without other 00:22:33.849 --> 00:22:36.509 strong guardrails in your prompt or tool configuration 00:22:36.509 --> 00:22:38.990 can lead to the agent getting stuck in infinite 00:22:38.990 --> 00:22:41.390 loops, just calling tools over and over. Oh, 00:22:41.490 --> 00:22:45.170 yeah. And racking up incredibly high costs very, 00:22:45.289 --> 00:22:47.730 very fast. It's a critical safety limit. You 00:22:47.730 --> 00:22:49.910 got to be careful with that one. Got it. Max 00:22:49.910 --> 00:22:52.869 iterations is a guardrail against infinite loops 00:22:52.869 --> 00:22:55.869 and uncontrolled spending. Anything else in that 00:22:55.869 --> 00:22:58.069 node we should know about? Return intermediate 00:22:58.069 --> 00:23:01.119 steps. This is super useful during development 00:23:01.119 --> 00:23:03.740 and debugging. How so? It shows you the agent's 00:23:03.740 --> 00:23:06.859 entire thought process. Every time it thinks, 00:23:07.000 --> 00:23:09.160 every tool it considers, every tool it calls, 00:23:09.420 --> 00:23:12.519 the results it got back, how it decided its next 00:23:12.519 --> 00:23:15.920 step. It's invaluable for understanding why an 00:23:15.920 --> 00:23:18.079 agent did what it did and troubleshooting problems. 00:23:18.380 --> 00:23:21.309 But you turn that off in production. For performance, 00:23:21.410 --> 00:23:23.769 maybe. Yes, definitely. It makes the output much 00:23:23.769 --> 00:23:25.650 cleaner for the rest of your workflow once you've 00:23:25.650 --> 00:23:27.910 got the agent working reliably. You don't need 00:23:27.910 --> 00:23:30.349 all that internal monologue clogging up the final 00:23:30.349 --> 00:23:32.349 result. Makes sense. And finally, automatically 00:23:32.349 --> 00:23:35.150 pass through binary images. That sounds specific. 00:23:35.490 --> 00:23:37.450 Yeah, you just want to enable that. If you're 00:23:37.450 --> 00:23:39.269 building multi -model workflows that involve 00:23:39.269 --> 00:23:41.890 images, for example, if you're sending an image 00:23:41.890 --> 00:23:45.809 to, say, GPT -4 for analysis, it just ensures 00:23:45.809 --> 00:23:48.349 the image data is handled correctly by the node 00:23:48.349 --> 00:23:51.559 and passed to the model. Simple toggle, but important 00:23:51.559 --> 00:23:54.279 for vision tasks. Okay, well, we covered all 00:23:54.279 --> 00:23:57.539 seven areas. Choosing the right brain or routing. 00:23:58.019 --> 00:24:00.960 Controlling creativity and reliability, TempTotP. 00:24:01.400 --> 00:24:04.140 Scalable memory with a database. Configuring 00:24:04.140 --> 00:24:07.000 tools intelligently. Predefining. Writing professional 00:24:07.000 --> 00:24:09.900 structured prompts. Using output parsers to guarantee 00:24:09.900 --> 00:24:12.720 format. And setting agent execution limits like 00:24:12.720 --> 00:24:15.740 max iterations. That is... That is so much more 00:24:15.740 --> 00:24:18.000 than just writing a prompt and, you know, hitting 00:24:18.000 --> 00:24:20.220 go. It really is. And the source material really 00:24:20.220 --> 00:24:22.559 nails the impact this has on the results you 00:24:22.559 --> 00:24:24.500 get. It's night and day. Yeah, let's talk about 00:24:24.500 --> 00:24:26.200 that. What's the real world difference between 00:24:26.200 --> 00:24:29.099 building an agent the amateur way versus applying 00:24:29.099 --> 00:24:30.880 these principles? What does it actually look 00:24:30.880 --> 00:24:32.400 like? Well, before applying these professional 00:24:32.400 --> 00:24:34.779 configurations, your agents work sometimes, right? 00:24:34.940 --> 00:24:37.859 Yeah, sometimes. Maybe. The outputs are unpredictable. 00:24:38.319 --> 00:24:40.819 They constantly break downstream workflows. Your 00:24:40.819 --> 00:24:42.720 costs are probably higher than they need to be 00:24:42.720 --> 00:24:44.579 because you're using the wrong models. And you 00:24:44.579 --> 00:24:47.900 spend hours manually debugging failures. It's 00:24:47.900 --> 00:24:51.200 honestly just an unreliable toy. Yeah, that sounds 00:24:51.200 --> 00:24:54.240 painfully familiar to, I think, anyone who's 00:24:54.240 --> 00:24:56.299 tried this at any scale. Yeah. That debugging 00:24:56.299 --> 00:24:59.099 time is killer. Exactly. After applying professional 00:24:59.099 --> 00:25:02.380 configuration, you get consistent... predictable 00:25:02.380 --> 00:25:05.640 behavior structured outputs that integrate seamlessly 00:25:05.640 --> 00:25:08.539 into the rest of your automation optimized costs 00:25:08.539 --> 00:25:10.579 because you're dynamically routing to the right 00:25:10.579 --> 00:25:13.680 model and resilient self -correcting systems 00:25:13.680 --> 00:25:16.299 you can actually depend on it transforms from 00:25:16.299 --> 00:25:19.680 a toy into a production asset So it is not just 00:25:19.680 --> 00:25:21.720 a little bit better quality output. It's about 00:25:21.720 --> 00:25:23.420 building something that's actually scalable, 00:25:23.640 --> 00:25:26.680 reliable, and can be the backbone for real business 00:25:26.680 --> 00:25:29.400 processes, something you can trust. Precisely. 00:25:29.400 --> 00:25:31.480 It's moving beyond the demo phase and building 00:25:31.480 --> 00:25:34.299 automation that can fundamentally change how 00:25:34.299 --> 00:25:35.980 you operate. It's serious business automation. 00:25:36.440 --> 00:25:39.119 This deep dive really shifted how I think about 00:25:39.119 --> 00:25:43.319 building these. It's not magic or just hoping 00:25:43.319 --> 00:25:46.869 the AI is smart enough. It's architecture. It's 00:25:46.869 --> 00:25:49.420 engineering. It's deliberate configuration. That's 00:25:49.420 --> 00:25:51.640 the core message the source wants to hammer home, 00:25:51.799 --> 00:25:54.940 I think. Stop treating NEs and AI agents like 00:25:54.940 --> 00:25:57.160 black boxes where you just throw a prompt in 00:25:57.160 --> 00:25:59.740 and hope. Start configuring them like the sophisticated, 00:26:00.079 --> 00:26:03.079 tunable tools they are. The competitive advantage 00:26:03.079 --> 00:26:05.319 isn't having some secret model nobody else has. 00:26:05.440 --> 00:26:07.960 It's how you configure, how you constrain, and 00:26:07.960 --> 00:26:09.900 how you architect them. That's where the value 00:26:09.900 --> 00:26:11.900 is. Yeah, it's like the slightly less glamorous 00:26:11.900 --> 00:26:14.160 stuff, the plumbing and the wiring. It actually 00:26:14.160 --> 00:26:17.140 makes the cool, visible stuff work reliably. 00:26:17.380 --> 00:26:19.180 The behind -the -scenes tuning. Totally. The 00:26:19.180 --> 00:26:21.779 companies and teams who are building truly reliable, 00:26:21.940 --> 00:26:24.519 innovative AI automation, they are absolutely 00:26:24.519 --> 00:26:26.680 using these kinds of professional configurations 00:26:26.680 --> 00:26:29.599 that, frankly, most online tutorials just never 00:26:29.599 --> 00:26:32.680 cover. This knowledge right here is the competitive 00:26:32.680 --> 00:26:35.480 advantage. That's a powerful takeaway. OK, so 00:26:35.480 --> 00:26:38.539 for you listening, maybe stop seeing your NNN 00:26:38.539 --> 00:26:41.380 AI agents as just, you know, toys you're hoping 00:26:41.380 --> 00:26:43.900 will perform well. Start seeing them as systems 00:26:43.900 --> 00:26:46.400 you can design, control and make dependable. 00:26:46.730 --> 00:26:49.109 And maybe think about this. What specific part 00:26:49.109 --> 00:26:51.529 of your current workflow is maybe relying on 00:26:51.529 --> 00:26:54.549 an unpredictable AI toy right now? Where are 00:26:54.549 --> 00:26:57.630 you just crossing your fingers? Hmm. Yeah. I 00:26:57.630 --> 00:27:00.059 can think of a few places in my own setups. How 00:27:00.059 --> 00:27:02.759 would applying even just one or two of these 00:27:02.759 --> 00:27:05.660 principles, maybe better model selection using 00:27:05.660 --> 00:27:08.400 a router, or dialing in temperature and top P 00:27:08.400 --> 00:27:11.000 correctly, or implementing structured output 00:27:11.000 --> 00:27:14.240 parsers, transform it from an unreliable mess 00:27:14.240 --> 00:27:16.859 into a dependable production grade system? Yeah. 00:27:16.900 --> 00:27:18.740 And how would that fundamental shift, building 00:27:18.740 --> 00:27:21.420 that reliability, change what you can even attempt 00:27:21.420 --> 00:27:23.920 to build next? What new possibilities open up 00:27:23.920 --> 00:27:25.509 when you know it's going to work? That's the 00:27:25.509 --> 00:27:27.750 provocative question, right? What becomes possible 00:27:27.750 --> 00:27:29.650 when the AI tool is truly reliable?