WEBVTT 00:00:00.000 --> 00:00:05.299 Imagine your business running itself, tasks handled, 00:00:05.660 --> 00:00:09.419 needs anticipated, problems just solved by intelligent 00:00:09.419 --> 00:00:12.759 digital teammates. Sounds like a dream, right? 00:00:13.039 --> 00:00:15.400 Well, maybe not anymore. This isn't really just 00:00:15.400 --> 00:00:17.579 sci -fi, you know, or some far -off promise. 00:00:18.480 --> 00:00:21.359 AI agents are fundamentally reshaping how enterprise 00:00:21.359 --> 00:00:24.059 works. Right now. Right now. Changing the whole 00:00:24.059 --> 00:00:26.879 idea of a workforce. Welcome to the Deep Dive. 00:00:27.019 --> 00:00:29.519 Today, we're taking a really close look at AI 00:00:29.519 --> 00:00:32.880 agents, these eponymous systems that are genuinely 00:00:32.880 --> 00:00:35.460 shaping the businesses of tomorrow. And our mission 00:00:35.460 --> 00:00:37.920 today, really, is to bridge that what -to -how 00:00:37.920 --> 00:00:40.740 gap for you will impact what AI agents truly 00:00:40.740 --> 00:00:44.140 are, why they represent such a monumental shift, 00:00:44.619 --> 00:00:46.479 where the biggest opportunities are, how you 00:00:46.479 --> 00:00:48.859 can actually identify and start building them, 00:00:48.859 --> 00:00:50.799 and also touch on some crucial ethical things 00:00:50.799 --> 00:00:52.840 we need to think about. Think of it as your strategic 00:00:52.840 --> 00:00:55.740 roadmap, a guide to this new frontier. exactly 00:00:55.740 --> 00:00:57.899 okay so let's start with the basics then what 00:00:57.899 --> 00:01:00.619 exactly is an AI agent if we boil it down simply 00:01:00.619 --> 00:01:03.780 put it's an autonomous system it can perceive 00:01:03.780 --> 00:01:07.200 its environment make its own decisions based 00:01:07.200 --> 00:01:09.719 on that perception and then carry out multi -step 00:01:09.719 --> 00:01:12.200 actions to hit a specific goal so we're not talking 00:01:12.200 --> 00:01:15.079 about a simple chat bot here no definitely not 00:01:15.079 --> 00:01:18.780 think of it more like a cognitive digital workforce 00:01:19.579 --> 00:01:21.680 And what's really fascinating is how they differ 00:01:21.680 --> 00:01:24.500 from, say, traditional software. Right. Traditional 00:01:24.500 --> 00:01:27.879 software is all about deterministic pre -programmed 00:01:27.879 --> 00:01:31.040 rules. It's rigid. It follows a script. OK. AI 00:01:31.040 --> 00:01:33.700 agents, though, they leverage things like large 00:01:33.700 --> 00:01:36.260 language models, LLMs, and other AI techniques 00:01:36.260 --> 00:01:39.060 to actually reason. They learn from feedback. 00:01:40.219 --> 00:01:42.840 And crucially, they adapt to new situations they 00:01:42.840 --> 00:01:44.700 haven't seen before. That's the key difference. 00:01:45.079 --> 00:01:47.260 It really is. It means they can operate 24 -7. 00:01:47.319 --> 00:01:49.659 They can scale on demand in ways humans just 00:01:49.659 --> 00:01:53.019 can't and deliver efficiency that's, well... 00:01:52.909 --> 00:01:54.790 pretty much unattainable otherwise. Yeah. It's 00:01:54.790 --> 00:01:57.870 a fundamental shift from just automation to genuine 00:01:57.870 --> 00:01:59.650 autonomy. And the numbers, I mean, they really 00:01:59.650 --> 00:02:01.129 back that up. They're quite dramatic. They are. 00:02:01.310 --> 00:02:03.950 Industry analysis projects the global AI agent 00:02:03.950 --> 00:02:07.189 market to surge from about $5 billion this year, 00:02:07.329 --> 00:02:12.969 2024, to over $47 billion by 2030. Wow. That's 00:02:12.969 --> 00:02:15.610 nearly a tenfold increase in just six years. 00:02:15.849 --> 00:02:18.050 This isn't just incremental growth. It feels 00:02:18.050 --> 00:02:20.469 like a fundamental reallocation of enterprise 00:02:20.469 --> 00:02:22.509 capital. Yeah, it signals that businesses aren't 00:02:22.509 --> 00:02:24.930 just experimenting anymore. They're making agents 00:02:24.930 --> 00:02:27.509 core to their strategy. And companies are already 00:02:27.509 --> 00:02:29.909 reporting huge results. Yeah. We're seeing things 00:02:29.909 --> 00:02:32.750 like 40, 80 percent operational cost reduction. 00:02:32.789 --> 00:02:35.650 This is massive. And even two to five times productivity 00:02:35.650 --> 00:02:38.430 gains where these agents are deployed. That 40, 00:02:38.430 --> 00:02:40.969 80 percent cost cut. I mean, that tells you entire 00:02:40.969 --> 00:02:43.330 workflows are being digitally optimized, not 00:02:43.330 --> 00:02:45.680 just slightly improved. It creates a real competitive 00:02:45.680 --> 00:02:48.139 gap, doesn't it, between the early adopters and 00:02:48.139 --> 00:02:50.520 everyone else? Definitely. But OK, let's clarify. 00:02:50.599 --> 00:02:52.580 What's the core difference, then, between an 00:02:52.580 --> 00:02:55.680 AI agent and just regular automation? The key 00:02:55.680 --> 00:02:58.199 thing is reasoning and adapting. Agents reason 00:02:58.199 --> 00:03:00.759 and adapt. Unlike static, rule -based automation, 00:03:01.039 --> 00:03:03.900 they can think. Got it. Adapting versus just 00:03:03.900 --> 00:03:07.180 following rules. So to really jump on this massive 00:03:07.180 --> 00:03:09.139 wave of change, we need to understand the dynamics, 00:03:09.219 --> 00:03:11.759 right? You mentioned a strategic roadmap, kind 00:03:11.759 --> 00:03:13.860 of like we saw with software as a service. Exactly. 00:03:14.770 --> 00:03:17.009 History gives us a bit of a blueprint here, if 00:03:17.009 --> 00:03:19.310 we look closely. We can think about it in waves. 00:03:19.490 --> 00:03:21.650 OK, waves. I like that. So wave one. Wave one 00:03:21.650 --> 00:03:23.930 was the obvious stuff that consumer applications 00:03:23.930 --> 00:03:28.069 think. Simple meeting summarizers, or maybe those 00:03:28.069 --> 00:03:31.129 email drafters. Low hanging fruit. Yeah, automating 00:03:31.129 --> 00:03:33.490 pretty straightforward digital tasks. But here's 00:03:33.490 --> 00:03:35.349 the interesting part, and the warning maybe. 00:03:36.210 --> 00:03:39.789 That space. is now completely commoditized. Right. 00:03:39.889 --> 00:03:43.069 Because the tech giants, Google, Microsoft, they 00:03:43.069 --> 00:03:45.770 just embed these features into their main products. 00:03:46.050 --> 00:03:49.069 Exactly. So if your AI agent idea can be easily 00:03:49.069 --> 00:03:51.789 copied as just a feature, well, your strategic 00:03:51.789 --> 00:03:53.889 mode is pretty shallow. It's dangerous territory. 00:03:54.110 --> 00:03:56.129 You know, honestly, I still sometimes wrestle 00:03:56.129 --> 00:03:59.490 with prompt drift myself. That's when an AI kind 00:03:59.490 --> 00:04:01.889 of subtly deviates from what you want it to do 00:04:01.889 --> 00:04:04.830 over time, even in those supposedly simple applications. 00:04:05.270 --> 00:04:07.810 So imagine the complexity in these bigger enterprise 00:04:07.810 --> 00:04:10.189 systems. Oh, absolutely. It's a real challenge, 00:04:10.770 --> 00:04:13.930 which brings us to wave two. The non -obvious 00:04:13.930 --> 00:04:16.709 consumer applications, these are emerging right 00:04:16.709 --> 00:04:19.870 now. They're creating entirely new consumer behaviors, 00:04:20.069 --> 00:04:22.230 kind of like Airbnb did for finding a place to 00:04:22.230 --> 00:04:24.709 stay or TikTok did for video. Things we didn't 00:04:24.709 --> 00:04:27.350 know we needed. Sort of, yeah. Yeah. And what's 00:04:27.350 --> 00:04:29.389 interesting is the big corporations are usually 00:04:29.389 --> 00:04:32.189 slow to jump into these because they seem unproven, 00:04:32.509 --> 00:04:37.149 risky. But for a bold entrepreneur, this wave 00:04:37.389 --> 00:04:40.449 has the potential to create category -defining 00:04:40.449 --> 00:04:42.910 companies. We're talking valuations in the tens, 00:04:43.189 --> 00:04:45.709 maybe hundreds of billions. OK, big potential 00:04:45.709 --> 00:04:49.170 there. And then wave three. Wave three. Vertical 00:04:49.170 --> 00:04:52.230 AI agents. Now, this, I genuinely believe, is 00:04:52.230 --> 00:04:54.870 the trillion -dollar opportunity. No exaggeration. 00:04:55.009 --> 00:04:56.910 A trillion? Wow, what does that mean, vertical? 00:04:57.110 --> 00:04:59.310 It means hyper -specialized solutions, tailored 00:04:59.310 --> 00:05:02.750 for unique, complex industry workflows. Imagine 00:05:02.750 --> 00:05:05.110 an AI agent built just for optimizing logistics 00:05:05.110 --> 00:05:08.170 networks. OK. Or one dedicated to ensuring pharmaceutical 00:05:08.170 --> 00:05:11.209 compliance. Or maybe an agent that intelligently 00:05:11.209 --> 00:05:13.629 optimizes crop yields in agriculture, pulling 00:05:13.629 --> 00:05:16.269 in weather data, soil data. Super specific. Exactly. 00:05:16.790 --> 00:05:19.470 And this category is a gold mine for a few key 00:05:19.470 --> 00:05:22.569 reasons. First, the total addressable market, 00:05:22.730 --> 00:05:25.689 the TAM, is massive. Right. You're not just going 00:05:25.689 --> 00:05:28.230 after a company's software budget. You're competing 00:05:28.230 --> 00:05:31.129 for that plus the payroll of the human teams 00:05:31.129 --> 00:05:33.610 doing that work now. The value proposition changes 00:05:33.610 --> 00:05:36.670 completely. From better software to a fully managed 00:05:36.670 --> 00:05:39.290 digital workforce. Revisedly. Which massively 00:05:39.290 --> 00:05:42.089 expands your potential revenue. Second, there's 00:05:42.089 --> 00:05:44.410 much lower competition from big tech. Why's that? 00:05:44.829 --> 00:05:46.970 Well, a hundred million dollar niche might be 00:05:46.970 --> 00:05:49.389 huge for a startup, right? But for a trillion 00:05:49.389 --> 00:05:51.290 dollar company, it's almost a rounding error. 00:05:51.370 --> 00:05:54.709 OK. Plus, they often lack the deep, deep domain 00:05:54.709 --> 00:05:57.110 expertise you need to build these really specialized 00:05:57.110 --> 00:05:59.759 solutions. This creates a kind of protected space 00:05:59.759 --> 00:06:02.759 for innovators. A defensible moat. Exactly. Built 00:06:02.759 --> 00:06:05.180 on specialized knowledge, which gives you a higher 00:06:05.180 --> 00:06:07.240 probability of success. So looking at those three 00:06:07.240 --> 00:06:09.759 ways, which one offers the clearest path, maybe 00:06:09.759 --> 00:06:12.040 the most accessible opportunity for someone starting 00:06:12.040 --> 00:06:15.360 out today? I'd say vertical AI agents offer the 00:06:15.360 --> 00:06:17.459 largest, most accessible opportunity for new 00:06:17.459 --> 00:06:20.839 builders. Less noise, clearer value. Yeah, I 00:06:20.839 --> 00:06:22.360 agree. That's where the smart money seems to 00:06:22.360 --> 00:06:24.670 be focusing. Okay, so understanding the theory 00:06:24.670 --> 00:06:27.350 of these waves is helpful, but finding a viable 00:06:27.350 --> 00:06:29.889 idea for one of these vertical agents, that seems 00:06:29.889 --> 00:06:32.649 like the real challenge. It is. So how do we 00:06:32.649 --> 00:06:36.189 systematically approach finding and, maybe more 00:06:36.189 --> 00:06:38.670 importantly, validating that opportunity? It 00:06:38.670 --> 00:06:41.189 really starts with knowing your ground. Step 00:06:41.189 --> 00:06:45.430 one, choose your domain wisely. Begin with what 00:06:45.430 --> 00:06:48.470 you already know, your insider knowledge. Your 00:06:48.470 --> 00:06:51.800 current job. A past career. Even a deeply understood 00:06:51.800 --> 00:06:54.639 hobby, potentially. Your nuanced understanding 00:06:54.639 --> 00:06:57.040 of the real -world problem that may be those 00:06:57.040 --> 00:06:59.980 hidden inefficiencies in a specific field, that's 00:06:59.980 --> 00:07:02.920 gold. So lean into your own experience. Absolutely. 00:07:03.439 --> 00:07:05.660 Promising general areas include things like logistics, 00:07:05.980 --> 00:07:08.160 legal tech, financial services, manufacturing, 00:07:08.720 --> 00:07:11.040 e -commerce, agriculture. But the best opportunities 00:07:11.040 --> 00:07:13.959 are often where you have direct personal experience 00:07:13.959 --> 00:07:16.379 with the pain points. Okay, that makes sense. 00:07:16.500 --> 00:07:19.629 Step two. Step two. Identify repetitive, high 00:07:19.629 --> 00:07:22.110 -value workflows. You're basically hunting for 00:07:22.110 --> 00:07:24.110 processes that are done over and over, but maybe 00:07:24.110 --> 00:07:26.569 have lots of variations. Stuff that rules -based 00:07:26.569 --> 00:07:29.569 systems choke on. Exactly. Look for tasks that 00:07:29.569 --> 00:07:32.569 are high volume, super time -intensive, or where 00:07:32.569 --> 00:07:36.189 mistakes are really costly. And critically, find 00:07:36.189 --> 00:07:39.110 workflows where a human is currently acting as 00:07:39.110 --> 00:07:42.029 the glue between different software tools. Can 00:07:42.029 --> 00:07:43.990 you give an example of that glue work? Sure. 00:07:44.170 --> 00:07:46.230 Think about a freight forwarder. They might be 00:07:46.230 --> 00:07:48.430 constantly pulling data from different shipping 00:07:48.430 --> 00:07:50.649 line websites, then re -entering it into their 00:07:50.649 --> 00:07:53.350 main system, then updating clients by email, 00:07:53.689 --> 00:07:55.589 then generating customs paperwork from another 00:07:55.589 --> 00:07:59.209 application. Ah, OK. Lots of manual copying and 00:07:59.209 --> 00:08:01.670 pasting and translating between systems. Precisely. 00:08:01.670 --> 00:08:03.949 That's a perfect target for an AI agent that 00:08:03.949 --> 00:08:06.490 can orchestrate all of that. Right. So step three, 00:08:06.790 --> 00:08:09.029 then, must be about leveraging the AI itself. 00:08:09.230 --> 00:08:12.569 Yep. Step three, leverage the unique advantages 00:08:12.569 --> 00:08:15.329 of AI agents. Design your solution around what 00:08:15.329 --> 00:08:18.949 they do best. Which is? 24 -7 operation handling 00:08:18.949 --> 00:08:21.470 global demands, different time zones, scalability 00:08:21.470 --> 00:08:23.689 processing thousands of things as easily as 10, 00:08:24.189 --> 00:08:26.569 consistency eliminating those human errors, and 00:08:26.569 --> 00:08:29.329 of course, significant cost efficiency. You're 00:08:29.329 --> 00:08:31.110 essentially offering a digital employee at a 00:08:31.110 --> 00:08:33.029 fraction of the cost. So it's about building 00:08:33.029 --> 00:08:35.850 a fundamentally better way, not just automating 00:08:35.850 --> 00:08:39.610 the old way. Exactly. And then step four, which 00:08:39.610 --> 00:08:42.649 is maybe the most crucial. Validation. Validate 00:08:42.649 --> 00:08:44.799 with extreme rigor. Seriously, before you write 00:08:44.799 --> 00:08:47.940 a single line of code, obsessively validate your 00:08:47.940 --> 00:08:50.840 idea. Interview potential customers. What kind 00:08:50.840 --> 00:08:53.059 of questions should you ask? Really probing ones. 00:08:53.480 --> 00:08:55.960 Walk me through the last time you handled whatever 00:08:55.960 --> 00:08:58.360 the target task is. Where did things get stuck? 00:08:58.419 --> 00:09:00.799 What frustrated you the most? Okay. Or maybe 00:09:00.799 --> 00:09:03.519 if you could wave a magic wand and just eliminate 00:09:03.519 --> 00:09:06.019 one part of your daily workload, what would it 00:09:06.019 --> 00:09:09.279 be and why? The goal is to find a problem that 00:09:09.279 --> 00:09:12.379 is so painful, customers will genuinely pay you 00:09:12.379 --> 00:09:15.000 to solve it. Ideally, they'll thank you for solving 00:09:15.000 --> 00:09:17.840 it. That's the dream. So the deepest understanding 00:09:17.840 --> 00:09:20.259 of the user's pain, that's the real key here, 00:09:20.279 --> 00:09:21.960 isn't it? To make sure you're building something 00:09:21.960 --> 00:09:24.220 people actually want. Absolutely. Deep empathy 00:09:24.220 --> 00:09:26.559 for the problem basically guarantees market demand. 00:09:26.679 --> 00:09:28.940 You're not guessing anymore. Couldn't agree more. 00:09:29.720 --> 00:09:31.980 Mid -roll sponsor read. Welcome back to the Deep 00:09:31.980 --> 00:09:34.610 Dive. We've talked about the what. and the why 00:09:34.610 --> 00:09:37.789 of AI agents, the strategic waves of opportunity, 00:09:38.090 --> 00:09:40.529 and how to find a good vertical idea. Right. 00:09:40.570 --> 00:09:42.450 So we've laid out the theory and the strategy. 00:09:42.789 --> 00:09:45.429 Now let's pivot to the practical side, the blueprint 00:09:45.429 --> 00:09:48.610 for actually building production -grade AI agents. 00:09:48.909 --> 00:09:51.110 Because having the idea is one thing, execution 00:09:51.110 --> 00:09:53.750 is another. Totally. Creating a really robust 00:09:53.750 --> 00:09:57.409 AI agent takes a mix of skills. It's multidisciplinary. 00:09:57.750 --> 00:10:00.370 We can anchor it around four crucial areas for 00:10:00.370 --> 00:10:02.690 pillars of competency. Okay, pillar one. Pillar 00:10:02.690 --> 00:10:05.409 one, advanced prompt engineering. This is really 00:10:05.409 --> 00:10:07.710 the art of instruction. It's about designing 00:10:07.710 --> 00:10:10.950 prompts that guide an LLM to perform accurately, 00:10:11.549 --> 00:10:13.669 reliably, consistently. We're talking way beyond 00:10:13.669 --> 00:10:16.289 just asking a simple question. Oh, yeah. An agent 00:10:16.289 --> 00:10:19.690 needs to handle complex, multi -step tasks, often 00:10:19.690 --> 00:10:21.929 from a single, very detailed prompt. It's what 00:10:21.929 --> 00:10:23.970 some call the structure flexibility paradox. 00:10:24.169 --> 00:10:25.889 Structure flexibility paradox. Explain that. 00:10:25.889 --> 00:10:28.370 You need to provide a rigid framework, a structure. 00:10:28.620 --> 00:10:31.620 for consistency. But you also need to allow the 00:10:31.620 --> 00:10:34.480 AI enough flexibility within that structure to 00:10:34.480 --> 00:10:37.279 reason and handle unexpected variations. So it's 00:10:37.279 --> 00:10:39.960 like giving super precise instructions to a smart 00:10:39.960 --> 00:10:42.919 intern. That's a great analogy. You define their 00:10:42.919 --> 00:10:45.559 exact role, give them all the necessary context, 00:10:45.980 --> 00:10:48.440 list the tools they can use, detail every single 00:10:48.440 --> 00:10:51.779 step, set clear boundaries, and specify the exact 00:10:51.779 --> 00:10:54.169 output format you need. The more rigorous the 00:10:54.169 --> 00:10:56.809 brief, the less room for error. Precisely. Less 00:10:56.809 --> 00:11:00.529 room for creative misinterpretation by the AI. 00:11:00.710 --> 00:11:04.169 OK, makes sense. Pillar two. Pillar two. Evaluation 00:11:04.169 --> 00:11:07.789 systems or evils. If prompt engineering is the 00:11:07.789 --> 00:11:10.370 art, evils are definitely the science. OK. These 00:11:10.370 --> 00:11:13.590 are automated tests. They systematically measure 00:11:13.590 --> 00:11:16.330 your agent's performance. Think about it. When 00:11:16.330 --> 00:11:20.029 an agent is live, running autonomously, You can't 00:11:20.029 --> 00:11:22.429 manually check every single thing it does. Right. 00:11:22.450 --> 00:11:24.610 That wouldn't scale. Not at all. You need systems 00:11:24.610 --> 00:11:26.830 constantly running maybe thousands of test scenarios 00:11:26.830 --> 00:11:29.149 to ensure it's reliable, especially as you make 00:11:29.149 --> 00:11:32.090 changes or the data it sees evolves. A robust 00:11:32.090 --> 00:11:34.649 evaluation suite is often a company's most valuable 00:11:34.649 --> 00:11:37.330 proprietary IP. Because it ensures trust and 00:11:37.330 --> 00:11:39.929 consistency. What kind of things do evils measure? 00:11:40.269 --> 00:11:43.500 Key things. Did it actually complete the task 00:11:43.500 --> 00:11:46.100 successfully? How good was the output quality? 00:11:46.559 --> 00:11:48.860 Did it use its tools correctly? How well did 00:11:48.860 --> 00:11:51.600 it recover from errors? And just how efficient 00:11:51.600 --> 00:11:54.700 was it? Okay, so evils are critical. Pillar three. 00:11:54.899 --> 00:11:58.539 Pillar three. Good old traditional software engineering. 00:11:59.240 --> 00:12:01.440 An AI agent isn't just a fancy prompt, it's a 00:12:01.440 --> 00:12:04.240 software product. Right. It needs a solid engineering 00:12:04.240 --> 00:12:06.850 foundation. That means designing for scalability 00:12:06.850 --> 00:12:09.450 right from the start, handling complex data flows 00:12:09.450 --> 00:12:12.169 properly, implementing really robust security. 00:12:12.350 --> 00:12:14.190 Security against things like prompt injection 00:12:14.190 --> 00:12:17.929 or data leaks. Exactly. And building good observability 00:12:17.929 --> 00:12:21.169 ways to monitor and debug the system when things 00:12:21.169 --> 00:12:23.970 go wrong. Wow. I mean, imagine scaling this to 00:12:23.970 --> 00:12:26.350 handle a billion queries a day. Yeah, that's 00:12:26.350 --> 00:12:28.450 where solid engineering really makes the difference 00:12:28.450 --> 00:12:30.970 between a cool demo and a resilient, reliable 00:12:30.970 --> 00:12:33.210 system. Absolutely. It's non -negotiable at scale. 00:12:33.789 --> 00:12:36.889 And finally, pillar four. Which is? Pillar four. 00:12:37.250 --> 00:12:39.690 Product and domain acumen. Technical brilliance 00:12:39.690 --> 00:12:42.110 alone doesn't cut it. It's meaningless if it 00:12:42.110 --> 00:12:44.110 doesn't solve a real world problem effectively. 00:12:44.429 --> 00:12:46.899 So understanding the user, the business. deeply. 00:12:47.440 --> 00:12:49.779 Understanding the user workflows, the business 00:12:49.779 --> 00:12:52.480 constraints, and crucially designing the human 00:12:52.480 --> 00:12:54.559 -in -the -loop aspect. Knowing when the agent 00:12:54.559 --> 00:12:57.059 needs to raise its hand and ask for help. Precisely. 00:12:57.139 --> 00:12:59.500 Knowing exactly when to escalate a problem to 00:12:59.500 --> 00:13:02.259 a human expert and designing that handoff smoothly. 00:13:02.820 --> 00:13:05.519 Making sure the agent is an assistant, not just 00:13:05.519 --> 00:13:08.259 operating in isolation. That's a critical product 00:13:08.259 --> 00:13:10.639 decision. Okay. Those four pillars make sense. 00:13:10.779 --> 00:13:13.700 Pumping, evils, engineering, and product domain 00:13:13.700 --> 00:13:16.419 sense. Looking at those four, which one do you 00:13:16.419 --> 00:13:19.460 think new teams, maybe folks coming more from 00:13:19.460 --> 00:13:22.379 the AI side than traditional software, tend to 00:13:22.379 --> 00:13:25.379 overlook or underestimate the most? I'd say evaluation 00:13:25.379 --> 00:13:28.240 systems are crucial but often neglected, leading 00:13:28.240 --> 00:13:30.639 to unreliable agents. People get excited about 00:13:30.639 --> 00:13:32.659 the prompting but forget to rigorously test. 00:13:32.830 --> 00:13:34.830 Right. You build something cool, but you can't 00:13:34.830 --> 00:13:37.269 actually prove it works reliably. And that stalls 00:13:37.269 --> 00:13:39.769 progress fast. OK, that leads nicely into the 00:13:39.769 --> 00:13:42.669 next point. Building these powerful agents comes 00:13:42.669 --> 00:13:46.570 with responsibility, right? Oh, absolutely. Neglecting 00:13:46.570 --> 00:13:49.029 the ethical considerations isn't just bad form. 00:13:49.409 --> 00:13:52.610 It's a huge business and reputational risk. We've 00:13:52.610 --> 00:13:55.950 seen systems perpetuate bias or cause real harm 00:13:55.950 --> 00:13:57.990 if not designed thoughtfully. So what are we 00:13:57.990 --> 00:14:00.029 talking about specifically? We're talking careful 00:14:00.029 --> 00:14:02.789 bias mitigation. in the data you use and the 00:14:02.789 --> 00:14:05.509 models themselves, designing systems that are 00:14:05.509 --> 00:14:08.429 auditable, that log decisions so you have transparency 00:14:08.429 --> 00:14:10.529 and explainability. So you can understand why 00:14:10.529 --> 00:14:13.429 it did what it did. Yes. And clear accountability 00:14:13.429 --> 00:14:15.769 structures for when mistakes inevitably happen. 00:14:16.110 --> 00:14:18.830 Robust data privacy measures, of course. And 00:14:18.830 --> 00:14:20.929 that human in the loop we mentioned. It's not 00:14:20.929 --> 00:14:23.210 just a feature for better performance. It's a 00:14:23.210 --> 00:14:25.149 governance mechanism. It's a critical governance 00:14:25.149 --> 00:14:27.950 mechanism. A safety net for accountability, for 00:14:27.950 --> 00:14:30.570 security, for catching things the AI might miss. 00:14:30.730 --> 00:14:33.570 This convergence of skills, the prompting, the 00:14:33.570 --> 00:14:36.710 evils, the engineering, the product sense, the 00:14:36.710 --> 00:14:38.830 ethical awareness, it feels like it's creating 00:14:38.830 --> 00:14:42.029 a whole new kind of role. It absolutely is. We're 00:14:42.029 --> 00:14:44.629 seeing the rise of the AI agent engineer. This 00:14:44.629 --> 00:14:47.429 person is a true hybrid. Blending software engineering, 00:14:47.730 --> 00:14:50.009 data science. And product management instincts, 00:14:50.529 --> 00:14:53.549 yeah. Demand for these individuals is just skyrocketing. 00:14:53.639 --> 00:14:56.480 Salaries are reflecting that, often ranging from, 00:14:56.480 --> 00:15:01.220 say, $95 ,000 up to over $270 ,000 for experienced 00:15:01.220 --> 00:15:04.080 people. Wow. But the opportunity isn't just direct 00:15:04.080 --> 00:15:06.960 employment. It extends to high -value consulting 00:15:06.960 --> 00:15:09.399 and definitely entrepreneurship building those 00:15:09.399 --> 00:15:11.460 vertical agents we talked about. So when you 00:15:11.460 --> 00:15:13.779 mentioned human in the loop, just to circle back, 00:15:14.100 --> 00:15:17.600 is it really more than just a feature to, say, 00:15:17.940 --> 00:15:20.769 improve the agent's accuracy? Yes, it's fundamentally 00:15:20.769 --> 00:15:23.350 a critical governance mechanism for accountability 00:15:23.350 --> 00:15:26.629 and safety. It integrates essential human oversight. 00:15:26.769 --> 00:15:29.590 Got it. Crucial distinction. OK, so for everyone 00:15:29.590 --> 00:15:31.350 listening who's feeling inspired and maybe ready 00:15:31.350 --> 00:15:33.169 to take the leap, let's give them a practical 00:15:33.169 --> 00:15:35.570 framework, a systematic approach to go from just 00:15:35.570 --> 00:15:37.809 an idea to actually implementing something. Right, 00:15:37.809 --> 00:15:39.629 let's break it down into maybe six deliberate 00:15:39.629 --> 00:15:43.750 phases. Phase one, deep immersion and observation. 00:15:44.070 --> 00:15:47.169 Meaning? Shadowing potential users, watching 00:15:47.169 --> 00:15:50.139 them work. meticulously documenting their daily 00:15:50.139 --> 00:15:52.379 frustrations, their workarounds, the inefficiencies, 00:15:52.720 --> 00:15:54.759 really live the problem. Okay, get boots on the 00:15:54.759 --> 00:15:58.100 ground. Phase two, workflow decomposition. Take 00:15:58.100 --> 00:16:00.279 the target task and break it down into really 00:16:00.279 --> 00:16:02.720 granular steps. Map out all the dependencies, 00:16:02.919 --> 00:16:05.019 the decision points. Understand the flow completely. 00:16:05.320 --> 00:16:07.720 Phase three, system design and human in the loop 00:16:07.720 --> 00:16:10.899 strategy. Start creating flow diagrams for how 00:16:10.899 --> 00:16:13.940 the agent will work. critically define those 00:16:13.940 --> 00:16:15.840 human escalation paths right from the start. 00:16:15.919 --> 00:16:18.860 When does the human step in? Phase four. Prototyping 00:16:18.860 --> 00:16:22.000 the happy path. Build a minimal viable agent 00:16:22.000 --> 00:16:24.740 but focus initially on the most common ideal 00:16:24.740 --> 00:16:27.340 scenario. Get that working first. Don't try to 00:16:27.340 --> 00:16:30.799 boil the ocean immediately. Exactly. Phase five. 00:16:31.100 --> 00:16:33.480 Rigorous evaluation and iteration. Build those 00:16:33.480 --> 00:16:35.500 evaluation suites in parallel with your agent 00:16:35.500 --> 00:16:37.539 development. Don't wait till the end. Test as 00:16:37.539 --> 00:16:40.159 you go. Constantly. And refine based on the data 00:16:40.159 --> 00:16:41.980 from those tests, not just your gut feeling. 00:16:42.639 --> 00:16:45.480 And finally, phase six, gradual scaling. OK. 00:16:45.740 --> 00:16:48.399 Slowly expand the agent's capabilities. Carefully 00:16:48.399 --> 00:16:50.799 onboard your first beta users. Get that real 00:16:50.799 --> 00:16:53.799 world feedback and keep refining. Is it absolutely 00:16:53.799 --> 00:16:55.779 essential to follow these steps in this exact 00:16:55.779 --> 00:16:58.340 order or is there some flexibility? I'd argue 00:16:58.340 --> 00:17:00.899 yes. Systematic validation and iteration are 00:17:00.899 --> 00:17:04.279 really key for building robust, reliable solutions. 00:17:04.839 --> 00:17:07.420 Skipping steps usually leads to trouble later. 00:17:07.619 --> 00:17:10.220 Right. Discipline pays off. So let's recap the 00:17:10.220 --> 00:17:14.039 big idea here. The AI agent revolution. It isn't 00:17:14.039 --> 00:17:15.779 science fiction. It's not tomorrow. It's happening 00:17:15.779 --> 00:17:19.420 right now. It's reshaping industries. Yeah, and 00:17:19.420 --> 00:17:23.039 the true lasting value won't be captured by those 00:17:23.039 --> 00:17:26.279 building general all purpose tools. It'll be 00:17:26.279 --> 00:17:28.559 captured by those who apply these powerful tools 00:17:28.559 --> 00:17:31.619 with surgical precision. Solving high stakes, 00:17:31.660 --> 00:17:34.279 specific problems in specific industries. Exactly. 00:17:34.440 --> 00:17:36.440 It's really about building that fully managed 00:17:36.440 --> 00:17:39.039 digital workforce we talked about, creating those 00:17:39.039 --> 00:17:42.059 defensible moats through deep domain expertise 00:17:42.059 --> 00:17:44.880 and potentially defining entirely new markets. 00:17:44.980 --> 00:17:46.980 It requires that blend of technical skill and 00:17:46.980 --> 00:17:49.319 business understanding. This really feels like 00:17:49.319 --> 00:17:51.619 the next major frontier in enterprise innovation. 00:17:51.920 --> 00:17:53.759 OK, so let's make this really concrete. Here's 00:17:53.759 --> 00:17:56.059 your call to action broken down into actionable 00:17:56.059 --> 00:17:59.079 steps for anyone listening. Let's do it. This 00:17:59.079 --> 00:18:02.119 week. Pick a vertical, one you genuinely understand, 00:18:02.559 --> 00:18:04.500 where you have some insider knowledge. Start 00:18:04.500 --> 00:18:06.339 playing around with advanced prompting frameworks. 00:18:06.480 --> 00:18:08.819 Use tools like Claude, OpenAI, whatever you have 00:18:08.819 --> 00:18:11.259 access to, and immerse yourself in the online 00:18:11.259 --> 00:18:12.859 communities where people are sharing what they're 00:18:12.859 --> 00:18:16.299 learning. Okay, dive in. This month. This month. 00:18:17.059 --> 00:18:20.900 Identify one high -value, repetitive workflow 00:18:20.900 --> 00:18:24.200 within that vertical you chose. Sketch out the 00:18:24.200 --> 00:18:25.960 agent's logic, how you think it should work. 00:18:26.500 --> 00:18:29.180 Then, go talk to at least five potential customers. 00:18:29.940 --> 00:18:33.019 Validate the pain point. Is it real? Is it painful 00:18:33.019 --> 00:18:35.119 enough they'd pay to fix it? Crucial validation 00:18:35.119 --> 00:18:37.839 step this quarter. This quarter. Build a basic 00:18:37.839 --> 00:18:40.000 prototype that happy path agent we discussed. 00:18:40.279 --> 00:18:42.660 But just as importantly, maybe even more importantly, 00:18:43.140 --> 00:18:45.420 create your first set of automated evaluation 00:18:45.420 --> 00:18:47.980 tests. Start measuring. Measure its performance 00:18:47.980 --> 00:18:50.440 rigorously. Refine it based on that data, not 00:18:50.440 --> 00:18:52.359 just on whether it looks like it's working. And 00:18:52.359 --> 00:18:54.900 this year... Big goal. Aim to have a production 00:18:54.900 --> 00:18:58.660 -ready agent actually serving real users. Focus 00:18:58.660 --> 00:19:01.180 intensely on building a sustainable business 00:19:01.180 --> 00:19:04.480 model around it. Deliver undeniable, measurable 00:19:04.480 --> 00:19:08.039 ROI to those first crucial customers. The companies 00:19:08.039 --> 00:19:10.079 that are going to define the next decade, they 00:19:10.079 --> 00:19:12.359 will be the ones that master autonomous systems. 00:19:12.460 --> 00:19:14.440 No doubt. They'll be built by individuals, maybe 00:19:14.440 --> 00:19:17.099 people listening right now, who bridge that gap 00:19:17.099 --> 00:19:21.380 between AI's incredible potential and the market's 00:19:21.380 --> 00:19:24.099 real pressing needs. The question isn't if these 00:19:24.099 --> 00:19:26.440 tools will change the world anymore. It's who 00:19:26.440 --> 00:19:29.400 will direct that change. Who will be the architects? 00:19:29.859 --> 00:19:32.299 Will you just be a consumer of this new technology 00:19:32.299 --> 00:19:34.880 or will you be one of its builders? The opportunity 00:19:34.880 --> 00:19:37.000 is definitely here. It's time to build. Well 00:19:37.000 --> 00:19:38.940 said. Thank you everyone for joining us on this 00:19:38.940 --> 00:19:41.059 deep dive. We really hope this knowledge empowers 00:19:41.059 --> 00:19:43.599 you to take that next step. We'll be back next 00:19:43.599 --> 00:19:46.299 time with another deep dive into the ideas shaping 00:19:46.299 --> 00:19:48.440 our future. Out to your own music.