WEBVTT 00:00:00.000 --> 00:00:02.459 You've probably used an AI, right? You asked 00:00:02.459 --> 00:00:04.679 it for something, maybe a quick summary, some 00:00:04.679 --> 00:00:07.219 code, a social media post. Yeah, and you get 00:00:07.219 --> 00:00:09.820 something back. It works, technically. But it's 00:00:09.820 --> 00:00:12.880 usually just... Yeah. Okay. Average. That average 00:00:12.880 --> 00:00:14.740 result. Well, believe it or not, that's often 00:00:14.740 --> 00:00:16.620 not the model's fault entirely. It's more about 00:00:16.620 --> 00:00:18.839 the architecture. Architecture. How so? We're 00:00:18.839 --> 00:00:21.539 still giving it, like, simple instructions when 00:00:21.539 --> 00:00:25.140 we should be thinking in terms of recipes, design 00:00:25.140 --> 00:00:28.530 patterns. Okay. Welcome to the Deep Dive. This 00:00:28.530 --> 00:00:31.309 is for you, the listener, who's ready to go beyond 00:00:31.309 --> 00:00:33.630 just basic prompting. You're looking to build 00:00:33.630 --> 00:00:36.950 real AI agency. Exactly. We all know the basic 00:00:36.950 --> 00:00:39.810 parts, right? The model, that's the brain, the 00:00:39.810 --> 00:00:43.549 tools, and those are the hands, and evals, the 00:00:43.549 --> 00:00:46.299 quality checks. Crucial stuff. But those are 00:00:46.299 --> 00:00:48.799 just the ingredients. Spot on. Today we're digging 00:00:48.799 --> 00:00:51.259 into the structure, the recipes. We're exploring 00:00:51.259 --> 00:00:54.740 four critical design patterns that take an AI 00:00:54.740 --> 00:00:57.060 agent from just following a command to running 00:00:57.060 --> 00:00:59.420 a really sophisticated autonomous workflow. So 00:00:59.420 --> 00:01:01.659 we're shifting the focus from what AI can do 00:01:01.659 --> 00:01:04.519 to how we make sure it does it well, reliably, 00:01:04.980 --> 00:01:08.519 every single time. Let's start with recipe one, 00:01:08.840 --> 00:01:12.430 reflection. This one feels like maybe the easiest 00:01:12.430 --> 00:01:14.450 entry point, something you can use right away, 00:01:14.670 --> 00:01:16.870 no fancy tools needed. Yeah, it's definitely 00:01:16.870 --> 00:01:19.909 powerful and accessible. The core idea is simple. 00:01:20.510 --> 00:01:22.890 Build in self -critique. Don't just take the 00:01:22.890 --> 00:01:25.989 AI's first answer. Okay. Instead, you force it 00:01:25.989 --> 00:01:29.250 to review its own work. critique it against specific 00:01:29.250 --> 00:01:31.090 rules before it even thinks about rewriting. 00:01:31.390 --> 00:01:33.730 Right. Because the usual way, the basic prompting, 00:01:33.870 --> 00:01:35.870 it's often too vague. We ask it to write something, 00:01:35.989 --> 00:01:37.790 then we just say... Is that good? Make it better. 00:01:37.989 --> 00:01:40.810 Exactly. And better could mean anything. So the 00:01:40.810 --> 00:01:43.209 AI just kind of fiddles with words. It often 00:01:43.209 --> 00:01:45.629 misses the real problem. Yeah. So the pro -level 00:01:45.629 --> 00:01:48.569 move here, the expert recipe, is using a specific 00:01:48.569 --> 00:01:51.180 structured rubric. A rubric. Like in school? 00:01:51.519 --> 00:01:53.920 Sort of, yeah. It forces a structured analysis. 00:01:54.140 --> 00:01:56.480 Let's say the AI writes a blog post. Instead 00:01:56.480 --> 00:01:58.980 of make it better, you give it, say, three things 00:01:58.980 --> 00:02:01.780 to grade itself on. Scale of one to five. OK. 00:02:02.060 --> 00:02:04.480 Grade the intro for clarity and hook one to five. 00:02:04.620 --> 00:02:06.900 Yep. Then maybe grade the examples used. How 00:02:06.900 --> 00:02:08.879 relevant are they? One to five. And the call 00:02:08.879 --> 00:02:11.919 to action, is it clear? One to five. Exactly. 00:02:11.979 --> 00:02:13.979 You're making the agent switch gears. It goes 00:02:13.979 --> 00:02:17.219 from just a fast writing mode. To thinking mode. 00:02:17.400 --> 00:02:20.360 Analysis mode. Right, structured analysis. And 00:02:20.360 --> 00:02:23.659 the trick often is using structured formats, 00:02:24.259 --> 00:02:27.819 things like XML tags or maybe JSON. You put the 00:02:27.819 --> 00:02:31.430 draft. Inside draft tags? Inside rubric tags? 00:02:31.550 --> 00:02:33.530 Why do those tags matter so much? Why not just 00:02:33.530 --> 00:02:35.909 bullet points in the prompt? It's about signaling, 00:02:35.930 --> 00:02:39.330 really. When the AI sees those specific tags, 00:02:39.449 --> 00:02:42.490 draft, rubric, it knows its job isn't creative 00:02:42.490 --> 00:02:45.889 writing anymore. Its job is now logical parsing, 00:02:45.969 --> 00:02:48.849 following strict rules. It has to look inside 00:02:48.849 --> 00:02:51.270 those tags, find the flaws according to the rubric 00:02:51.270 --> 00:02:53.370 before it can try again. It literally needs to 00:02:53.370 --> 00:02:56.050 figure out why it got a 2 out of 5 on clarity 00:02:56.050 --> 00:02:57.810 before rewriting. That makes a lot of sense. 00:02:57.870 --> 00:02:59.689 It's like putting up guardrails. Okay, but let's 00:02:59.689 --> 00:03:02.669 talk trade -offs. Doesn't this double the time 00:03:02.669 --> 00:03:04.870 and the cost? You're basically running it twice. 00:03:05.110 --> 00:03:07.349 It absolutely does increase latency and cost. 00:03:07.530 --> 00:03:09.389 Yeah, it's a multi -step process. Takes longer, 00:03:09.409 --> 00:03:11.969 uses more tokens, no question. Why do it? Because 00:03:11.969 --> 00:03:14.650 you're paying for quality assurance. It's a trade 00:03:14.650 --> 00:03:17.909 -off, yes, but the improvement you get. Going 00:03:17.909 --> 00:03:21.210 from a med draft to a really solid final piece, 00:03:21.430 --> 00:03:24.789 that quality jump usually far outweighs the extra 00:03:24.789 --> 00:03:27.349 cost or time. It's an investment then. and getting 00:03:27.349 --> 00:03:29.849 it right. Precisely. I have to admit, though, 00:03:30.409 --> 00:03:32.830 even knowing this, sometimes I still catch myself 00:03:32.830 --> 00:03:35.789 typing that lazy, make -it -better prompt. You 00:03:35.789 --> 00:03:38.770 know, before I stop, delete it and actually build 00:03:38.770 --> 00:03:40.669 the rubric. It's a habit that's kind of hard 00:03:40.669 --> 00:03:42.530 to break. That's honest, and I think a lot of 00:03:42.530 --> 00:03:44.530 people can relate. It takes discipline. But OK, 00:03:44.550 --> 00:03:46.830 if you had to boil down the immediate benefit 00:03:46.830 --> 00:03:49.750 of Recipe 1, what's the core shift for the user? 00:03:49.990 --> 00:03:53.030 Quality goes up instantly, because the AI has 00:03:53.030 --> 00:03:56.389 to judge itself against clear, strict rules. 00:03:56.539 --> 00:03:59.400 OK, so reflection helps the AI improve itself 00:03:59.400 --> 00:04:02.219 internally. Recipe two, tool use, is about giving 00:04:02.219 --> 00:04:04.199 it external powers. Giving it hands, you said 00:04:04.199 --> 00:04:06.439 earlier. Connecting it to the outside world. 00:04:06.800 --> 00:04:10.199 APIs, search databases. Exactly. And modern models, 00:04:10.219 --> 00:04:12.060 they're pretty smart. They often know they might 00:04:12.060 --> 00:04:14.319 need a tool. But just relying on that built -in 00:04:14.319 --> 00:04:16.339 knowledge, that zero -shot ability, it can get 00:04:16.339 --> 00:04:18.819 tripped up by complex questions. Yeah, like if 00:04:18.819 --> 00:04:20.540 you ask something with multiple parts. Right. 00:04:20.800 --> 00:04:22.720 You ask, is it going to be cold in Hanoi tomorrow? 00:04:24.069 --> 00:04:28.170 And based on that, should I pack a jacket and 00:04:28.170 --> 00:04:30.370 maybe an umbrella? Okay, that's a few steps. 00:04:30.449 --> 00:04:32.350 Yeah, and the AI might just check the temperature 00:04:32.350 --> 00:04:34.629 but completely forget the part about the umbrella. 00:04:34.769 --> 00:04:37.050 It gets confused. So the pro -level technique 00:04:37.050 --> 00:04:39.589 is using guiding examples. Few -shot learning, 00:04:39.670 --> 00:04:41.910 right? Exactly. We don't just tell it the tool 00:04:41.910 --> 00:04:44.990 exists. We show it how to think through specific 00:04:44.990 --> 00:04:47.410 problems using examples. Like teaching it the 00:04:47.410 --> 00:04:50.589 decision process. Precisely. If a user asks that 00:04:50.589 --> 00:04:52.689 two -part question, maybe... How much does this 00:04:52.689 --> 00:04:55.750 cost and is it in stock? The example teaches 00:04:55.750 --> 00:04:58.310 the AI an internal thought process. Like what? 00:04:58.509 --> 00:05:00.310 Like, OK, wait, I need the product ID before 00:05:00.310 --> 00:05:05.230 I can check price or stock. So step one, I must 00:05:05.230 --> 00:05:08.959 use the ask you for info action first. Ah, so 00:05:08.959 --> 00:05:10.779 the key isn't just using the tool. It's teaching 00:05:10.779 --> 00:05:13.019 the AI when to stop and ask for missing info. 00:05:13.139 --> 00:05:16.079 Yes, that dramatically cuts down on errors. On 00:05:16.079 --> 00:05:18.860 the AI just guessing or making stuff up. So teaching 00:05:18.860 --> 00:05:21.560 the context, the steps needed, makes the whole 00:05:21.560 --> 00:05:24.060 system way more predictable and less error -prone. 00:05:24.240 --> 00:05:27.660 Totally. It reduces errors drastically. Makes 00:05:27.660 --> 00:05:31.740 it reliable. Yeah. Whoa. Just imagine scaling 00:05:31.740 --> 00:05:34.939 that. A system built with these robust guiding 00:05:34.939 --> 00:05:39.129 examples. Handling, say, a billion complex financial 00:05:39.129 --> 00:05:42.449 queries a day. Wow. That level of reliable connection 00:05:42.449 --> 00:05:45.550 to real -world systems, that unlocks some serious 00:05:45.550 --> 00:05:48.269 business value. Truly powerful stuff. Okay, so 00:05:48.269 --> 00:05:52.209 tools give the AI hands. This next recipe, planning. 00:05:53.579 --> 00:05:55.759 That sounds like giving it foresight. Yeah, now 00:05:55.759 --> 00:05:57.680 we're getting into really autonomous territory. 00:05:58.040 --> 00:05:59.939 The user just gives the high -level goal, and 00:05:59.939 --> 00:06:02.120 the agent has to figure out the entire step -by 00:06:02.120 --> 00:06:04.639 -step plan itself. But basic planning often fails, 00:06:04.680 --> 00:06:06.600 right? It does, because the AI tends to jump 00:06:06.600 --> 00:06:08.579 the gun. You say... Plan a trip to Paris for 00:06:08.579 --> 00:06:10.240 three days. And it immediately starts looking 00:06:10.240 --> 00:06:13.139 for flights. Exactly. Flights, hotels. It skips 00:06:13.139 --> 00:06:14.839 the crucial first step, asking questions. What's 00:06:14.839 --> 00:06:16.839 the budget? When are you going? Who's even traveling? 00:06:16.920 --> 00:06:19.759 It just defaults to a generic plan. Right. Minimum 00:06:19.759 --> 00:06:21.779 effort solution. Yeah. So the better structure, 00:06:22.000 --> 00:06:25.519 often called React, reasoning and action. It 00:06:25.519 --> 00:06:28.459 forces a mandatory plan and critique cycle. Oh, 00:06:28.459 --> 00:06:30.680 cycle. Yeah. A rigid four -step process it must 00:06:30.680 --> 00:06:32.620 follow before it does anything. OK, what are 00:06:32.620 --> 00:06:35.639 the steps? Thought. Clearly define the user's 00:06:35.639 --> 00:06:39.160 ultimate goal. Step two. Initial plan. Write 00:06:39.160 --> 00:06:42.379 down the steps it thinks it needs. One. Two. 00:06:42.819 --> 00:06:45.459 Three. Makes sense. This is the core self -critique. 00:06:45.699 --> 00:06:48.600 The AI must ask itself. What information am I 00:06:48.600 --> 00:06:53.120 missing? Budget. Dates. Interests. Where could 00:06:53.120 --> 00:06:55.899 this plan go wrong? This is where it confronts 00:06:55.899 --> 00:06:58.079 its own ignorance. It forces it to see the gaps. 00:06:58.240 --> 00:07:02.180 Exactly. Then step four. Final plan. It takes 00:07:02.180 --> 00:07:05.500 the critique and fixes the plan. And here's the 00:07:05.500 --> 00:07:08.519 key. The very first step in that final plan, 00:07:08.740 --> 00:07:10.519 almost always it should be using that ask user 00:07:10.519 --> 00:07:12.839 for info tool we talked about. Ah, connecting 00:07:12.839 --> 00:07:15.300 back to tool use. Yep. This whole structure stops 00:07:15.300 --> 00:07:17.540 the AI from just rushing ahead vaguely. It makes 00:07:17.540 --> 00:07:20.160 sure it gets the necessary info first before 00:07:20.160 --> 00:07:22.259 wasting time or, you know, computing resources. 00:07:22.420 --> 00:07:24.720 So that mandatory self -critique phase forces 00:07:24.720 --> 00:07:27.279 it to gather info first. That one step prevents 00:07:27.279 --> 00:07:29.800 the rush to a bad generic solution. It builds 00:07:29.800 --> 00:07:31.639 quality in right from the start. Okay, that makes 00:07:31.639 --> 00:07:33.360 a lot of sense. We need to take a quick break. 00:07:33.420 --> 00:07:34.939 When we come back, we'll dive into the final 00:07:34.939 --> 00:07:36.920 most advanced recipe, multi -agent workflows. 00:07:37.279 --> 00:07:39.879 Sounds good. Mid -roll sponsor read content to 00:07:39.879 --> 00:07:42.759 be inserted by Rider. All right, we're back. 00:07:42.879 --> 00:07:45.639 We've covered reflection, tool use, and planning. 00:07:46.500 --> 00:07:48.360 Now for the final recipe, which you said is the 00:07:48.360 --> 00:07:51.740 most advanced, multi -agent workflow. Yeah, this 00:07:51.740 --> 00:07:55.199 is where we stop trying to make one single general 00:07:55.199 --> 00:07:58.540 AI do everything perfectly. Instead, we build 00:07:58.540 --> 00:08:02.019 a team. A team of specialized AI agents. OK, 00:08:02.019 --> 00:08:04.399 because the normal way, that's usually just us, 00:08:04.500 --> 00:08:06.959 right? Doing the glue work. Exactly. Manual glue 00:08:06.959 --> 00:08:10.459 work. You ask your research AI for market trends. 00:08:10.620 --> 00:08:13.199 You copy the text. You paste it into your writing 00:08:13.199 --> 00:08:15.680 AI to generate some ad copy. And you're the one 00:08:15.680 --> 00:08:17.839 connecting the dots, trying to keep the context 00:08:17.839 --> 00:08:20.139 straight between copy paste. Right. And context 00:08:20.139 --> 00:08:23.360 gets lost or muddied. It's inefficient. The pro 00:08:23.360 --> 00:08:25.920 approach is setting up clear specialized roles 00:08:25.920 --> 00:08:28.199 for each agent and really strict rules for how 00:08:28.199 --> 00:08:30.220 they hand off. information. Think of it like 00:08:30.220 --> 00:08:32.600 a hyper -specialized assembly line. An assembly 00:08:32.600 --> 00:08:35.320 line for AI. Pretty much. The key concept here 00:08:35.320 --> 00:08:37.419 is fighting something called function creep. 00:08:37.679 --> 00:08:39.840 Function creep? What's that? It's when a single 00:08:39.840 --> 00:08:42.000 agent starts getting asked to do jobs it wasn't 00:08:42.000 --> 00:08:45.259 really designed or optimized for. Its core quality 00:08:45.259 --> 00:08:47.200 slowly degrades because it's being stretched 00:08:47.200 --> 00:08:49.919 too thin. Specialization prevents that. Okay, 00:08:49.940 --> 00:08:51.899 let's use that ad campaign example you mentioned. 00:08:52.279 --> 00:08:54.980 So instead of one AI, we have three. Yeah, let's 00:08:54.980 --> 00:08:58.100 say three distinct roles. Agent one. We'll call 00:08:58.100 --> 00:09:02.600 him Data Dave. Data Dave. OK. His job is purely 00:09:02.600 --> 00:09:06.580 logical, factual, analyze market trends. And 00:09:06.580 --> 00:09:09.480 crucially, his output must be a very strict JSON 00:09:09.480 --> 00:09:12.879 object. Maybe detailing target audience, key 00:09:12.879 --> 00:09:15.879 trends, the core pain point. He only speaks data, 00:09:16.399 --> 00:09:19.399 no creativity allowed. Strict JSON. Got it. Then 00:09:19.399 --> 00:09:22.539 that JSON goes to? Creative Carla. Her role is 00:09:22.539 --> 00:09:25.720 all about emotion narrative. turning Dave's dry 00:09:25.720 --> 00:09:27.899 data into, say, three compelling ad options. 00:09:27.960 --> 00:09:29.740 Yeah, because she gets that strict JSON contract 00:09:29.740 --> 00:09:31.340 from Dave. She does exactly which pain points 00:09:31.340 --> 00:09:33.519 she needs to address. No guesswork. OK, makes 00:09:33.519 --> 00:09:36.279 sense. And agent three. Manager Mike. His role 00:09:36.279 --> 00:09:38.480 is practical, results focused. He takes Carla's 00:09:38.480 --> 00:09:40.779 creative ads, compares them against Dave's original 00:09:40.779 --> 00:09:42.480 data analysis. Checks if they actually match 00:09:42.480 --> 00:09:44.879 the research. Right. And then he picks the single 00:09:44.879 --> 00:09:47.779 best ad and explains why based on the data and 00:09:47.779 --> 00:09:51.110 the goal. OK, but. Couldn't you just use one 00:09:51.110 --> 00:09:54.330 really big powerful model like GPT -4 for all 00:09:54.330 --> 00:09:57.110 of this? Seems way less complicated than setting 00:09:57.110 --> 00:09:59.649 up and managing this whole AI team. It might 00:09:59.649 --> 00:10:02.230 seem simpler up front, yeah, but the quality 00:10:02.230 --> 00:10:04.889 difference can be huge. That specialization is 00:10:04.889 --> 00:10:07.649 key. Why though, if the big model is smart enough? 00:10:08.309 --> 00:10:10.070 Because of that function creep we talked about. 00:10:10.269 --> 00:10:13.049 Data Dave cannot write good ad copy. Creative 00:10:13.049 --> 00:10:16.129 Carla cannot do rigorous data analysis. Forcing 00:10:16.129 --> 00:10:18.649 them into specialized lanes maintains peak performance 00:10:18.649 --> 00:10:21.110 for their specific task. And the other crucial 00:10:21.110 --> 00:10:23.889 piece, that strict JSON format for handoffs, 00:10:24.289 --> 00:10:26.669 it acts like an unbreakable contract between 00:10:26.669 --> 00:10:30.490 them. It drastically reduces the risk of hallucination 00:10:30.490 --> 00:10:32.690 or misinterpretation as information moves down 00:10:32.690 --> 00:10:35.409 the line. Context is preserved perfectly. The 00:10:35.409 --> 00:10:37.950 quality boost comes from forcing specialization 00:10:37.950 --> 00:10:40.090 and using that strict data contract. between 00:10:40.090 --> 00:10:42.129 them. Exactly. Specialization stops function 00:10:42.129 --> 00:10:45.429 creep. Strict JSON ensures perfect context transfer. 00:10:45.610 --> 00:10:47.750 You get a better final product from the assembly 00:10:47.750 --> 00:10:50.450 line. Okay, so let's recap the big ideas. Four 00:10:50.450 --> 00:10:53.309 recipes for moving beyond basic prompts. Yep. 00:10:53.649 --> 00:10:56.070 First, reflection improving internal quality 00:10:56.070 --> 00:10:59.110 using a strict rubric for self -critique. Second, 00:10:59.789 --> 00:11:02.669 tool use giving the AI external capabilities 00:11:02.669 --> 00:11:05.490 but guiding it with few -shot examples so it 00:11:05.490 --> 00:11:08.440 knows when and how. Third, planning creating 00:11:08.440 --> 00:11:11.059 autonomy with that mandatory self -critique step 00:11:11.059 --> 00:11:13.419 to force information gathering before acting, 00:11:13.940 --> 00:11:16.059 using structures like React. And fourth, multi 00:11:16.059 --> 00:11:18.259 -agent building specialized teams that communicate 00:11:18.259 --> 00:11:21.419 via strict formats like JSON to maximize quality 00:11:21.419 --> 00:11:24.039 and avoid function creed. That's the progression. 00:11:24.320 --> 00:11:27.139 From simple internal checks to complex orchestrated 00:11:27.139 --> 00:11:29.539 teams. Now, you probably shouldn't try to build 00:11:29.539 --> 00:11:32.129 a complex multi -agent system tomorrow. Definitely 00:11:32.129 --> 00:11:34.769 not. My advice is start simple. Get comfortable 00:11:34.769 --> 00:11:37.330 with the earlier recipes first. Right. So here's 00:11:37.330 --> 00:11:39.169 some actionable homework for you, the listener. 00:11:39.529 --> 00:11:42.090 Try recipe one, reflection, today. Yeah, pick 00:11:42.090 --> 00:11:44.129 a simple task you do often, maybe writing email 00:11:44.129 --> 00:11:47.590 subject lines. OK. Ask your AI to write, say, 00:11:47.909 --> 00:11:50.610 three subject lines for an email. Then immediately 00:11:50.610 --> 00:11:53.309 after, give it a rubric. Tell it. Grade these 00:11:53.309 --> 00:11:56.210 subject lines from one to five on curiosity and 00:11:56.210 --> 00:11:58.289 one to five on urgency make it be strict and 00:11:58.289 --> 00:12:01.029 then the final step Demand new subject lines 00:12:01.029 --> 00:12:03.750 tell it now write three new ones that score five 00:12:03.750 --> 00:12:06.929 out of five on both curiosity and Urgency based 00:12:06.929 --> 00:12:09.450 on your own critique and just notice the difference 00:12:09.450 --> 00:12:12.490 between that first attempt and the refined reflected 00:12:12.490 --> 00:12:14.990 version you can be dramatic it really shows that 00:12:14.990 --> 00:12:17.879 if Just changing a prompt slightly, adding that 00:12:17.879 --> 00:12:20.179 reflection step can give you such a better result. 00:12:21.039 --> 00:12:23.120 Imagine the power when you start layering these 00:12:23.120 --> 00:12:26.259 patterns, tool use on top of reflection, planning, 00:12:26.279 --> 00:12:28.740 managing tool use. It compounds. So here's a 00:12:28.740 --> 00:12:30.259 final thought to leave you with, something to 00:12:30.259 --> 00:12:33.039 chew on. OK. If your AI agent, the one you use 00:12:33.039 --> 00:12:35.139 every day, if it always checked what critical 00:12:35.139 --> 00:12:36.799 information it was missing before it took any 00:12:36.799 --> 00:12:39.919 action. How would that one change, that architectural 00:12:39.919 --> 00:12:42.279 shift based on the planning recipe, how would 00:12:42.279 --> 00:12:44.700 that fundamentally change your daily workflow, 00:12:44.879 --> 00:12:47.559 your decision -making? That's a really interesting 00:12:47.559 --> 00:12:50.159 question to consider. How much guesswork would 00:12:50.159 --> 00:12:53.179 that eliminate? Something to think about. Definitely. 00:12:53.960 --> 00:12:55.980 A thought to carry with you until our next deep 00:12:55.980 --> 00:12:57.620 dive. Thanks for tuning in.