WEBVTT 00:00:00.000 --> 00:00:03.200 The days when you absolutely needed a computer 00:00:03.200 --> 00:00:05.440 science degree, you know, just to automate some 00:00:05.440 --> 00:00:08.119 basic tasks, those days are really fading. It's 00:00:08.119 --> 00:00:10.359 a pretty profound shift, actually. Today we're 00:00:10.359 --> 00:00:13.199 doing a deep dive into a platform that I think 00:00:13.199 --> 00:00:16.719 really embodies this change, the chat GPT agent 00:00:16.719 --> 00:00:19.399 builder. And let's pretty much anyone build these 00:00:19.399 --> 00:00:22.890 sophisticated AI helpers using, well, just. Visual 00:00:22.890 --> 00:00:24.949 blocks and an idea welcome. Yeah Our mission 00:00:24.949 --> 00:00:27.469 today is to really unpack this thing get to the 00:00:27.469 --> 00:00:30.429 core of how it works and the tool itself It's 00:00:30.429 --> 00:00:34.030 a platform for creating smart tasks specific 00:00:34.030 --> 00:00:36.350 AI assistance these agents without touching a 00:00:36.350 --> 00:00:38.929 line of code Okay, so just for clarity then define 00:00:38.929 --> 00:00:41.770 AI agent for us quickly What is it really think 00:00:41.770 --> 00:00:44.170 of it like a highly focused digital employee? 00:00:44.549 --> 00:00:46.570 Basically smart helpers designed to automate 00:00:46.570 --> 00:00:49.009 very specific tasks and they operate purely based 00:00:49.009 --> 00:00:51.399 on the rules you give them Perfect. So the plan 00:00:51.399 --> 00:00:55.159 today, first we'll look at the basic logic, the 00:00:55.159 --> 00:00:57.979 blueprint, using some examples. Then we'll actually 00:00:57.979 --> 00:00:59.979 walk through building an educational agent step 00:00:59.979 --> 00:01:02.759 by step, one with multiple paths. And finally, 00:01:03.520 --> 00:01:05.500 we'll look at how you connect these agents out 00:01:05.500 --> 00:01:10.019 to other apps, like Gmail or Shopify. All right, 00:01:10.019 --> 00:01:12.280 let's unpack that blueprint. The tool sounds 00:01:12.280 --> 00:01:15.269 really accessible, but you know. Accessibility 00:01:15.269 --> 00:01:18.290 isn't the same as it being easy to design something 00:01:18.290 --> 00:01:20.349 good. What's the actual biggest hurdle here? 00:01:20.750 --> 00:01:22.730 It's almost always conceptual, actually. I mean, 00:01:22.810 --> 00:01:24.650 the technical part. It's practically zero now. 00:01:25.030 --> 00:01:27.510 But your ability to clearly define the problem, 00:01:27.730 --> 00:01:30.049 the instructions you write, they have to be absolutely 00:01:30.049 --> 00:01:32.930 spot on, really specific for the agent to work 00:01:32.930 --> 00:01:36.299 reliably. It sounds incredibly powerful, but 00:01:36.299 --> 00:01:38.579 the mechanism you described, it's kind of like 00:01:38.579 --> 00:01:40.859 stacking Lego blocks almost. That's a perfect 00:01:40.859 --> 00:01:42.680 way to put it, yeah. If you can sketch out a 00:01:42.680 --> 00:01:44.739 simple flow chart, you can build an agent. Each 00:01:44.739 --> 00:01:47.239 block does one job. Maybe it collects data, searches 00:01:47.239 --> 00:01:49.739 the web, makes a decision based on a rule. You 00:01:49.739 --> 00:01:51.900 just connect them together to get the workflow 00:01:51.900 --> 00:01:54.640 you need. And getting started. People need to 00:01:54.640 --> 00:01:57.280 go into the OpenAI workspace, set up payment 00:01:57.280 --> 00:01:59.819 for usage first. That's right. Yeah, that's just 00:01:59.819 --> 00:02:02.060 to cover the computing resources the agent uses 00:02:02.060 --> 00:02:04.359 when it's running. Once you're in, you see the 00:02:04.359 --> 00:02:06.540 blocks on the left and this main canvas, the 00:02:06.540 --> 00:02:09.340 workspace where you arrange everything. OK, so 00:02:09.340 --> 00:02:13.020 if the tech barrier is basically gone and the 00:02:13.020 --> 00:02:15.639 main challenge is being clear in your thinking, 00:02:17.400 --> 00:02:19.580 what's the limiting factor then? Is it really 00:02:19.580 --> 00:02:22.259 just how well you understand your own process? 00:02:22.539 --> 00:02:24.939 Yeah, essentially. Your instructions must be 00:02:24.939 --> 00:02:26.819 absolutely clear for the agent to work well. 00:02:27.099 --> 00:02:30.080 That clarity piece leads nicely into agent logic. 00:02:30.659 --> 00:02:33.199 How these things actually make decisions. It 00:02:33.199 --> 00:02:35.280 feels like we're moving away from general chat 00:02:35.280 --> 00:02:37.810 towards something more structured. I like your 00:02:37.810 --> 00:02:40.569 idea of a complex agent being like a small, really 00:02:40.569 --> 00:02:42.509 specialized team. Exactly. Take the planning 00:02:42.509 --> 00:02:45.090 helper agent example they give. It kicks off 00:02:45.090 --> 00:02:46.990 with a start trigger that's just the moment the 00:02:46.990 --> 00:02:49.330 user submits their request, like ringing a doorbell. 00:02:49.710 --> 00:02:51.810 Then you've got the triage agent. The information 00:02:51.810 --> 00:02:53.830 gatherer, right. Its only job is grabbing all 00:02:53.830 --> 00:02:56.669 the key project details right away. Right. And 00:02:56.669 --> 00:02:58.789 then comes the condition check. This is kind 00:02:58.789 --> 00:03:01.590 of the smart bit. It asks a simple yes -no question. 00:03:01.949 --> 00:03:04.909 Do I have all the info I need? If it's yes, yes, 00:03:05.169 --> 00:03:07.789 okay, proceed to planning. If it's no, it might 00:03:07.789 --> 00:03:10.030 loop back or trigger another agent like a get 00:03:10.030 --> 00:03:12.710 data agent to ask for what's missing. So you 00:03:12.710 --> 00:03:15.430 see branching logic based on the data flow. The 00:03:15.430 --> 00:03:17.409 customer service agent example is another good 00:03:17.409 --> 00:03:19.449 one for branching. The first agent just figures 00:03:19.449 --> 00:03:21.229 out the intent. Is this person trying to return 00:03:21.229 --> 00:03:24.740 something? Cancel. Just get info. And that intent 00:03:24.740 --> 00:03:27.520 immediately dictates which specialized agent 00:03:27.520 --> 00:03:29.800 takes over. The return agent, the retention agent, 00:03:29.919 --> 00:03:32.280 whatever it is. But here's maybe a question. 00:03:32.620 --> 00:03:34.979 Doesn't using multiple specialized agents potentially 00:03:34.979 --> 00:03:38.439 add latency or cost compared to just one big, 00:03:38.719 --> 00:03:40.860 smarter model? Ah, that's the interesting trade 00:03:40.860 --> 00:03:43.000 -off, isn't it? By using these specialized agents, 00:03:43.360 --> 00:03:45.340 you can actually assign different types of AI 00:03:45.340 --> 00:03:47.460 models to different parts of the job. A simple 00:03:47.460 --> 00:03:49.379 sorting task doesn't need the most powerful, 00:03:49.400 --> 00:03:52.539 most expensive AI. Precisely. You use the big 00:03:52.539 --> 00:03:55.819 guns, maybe like the latest GPT model, for the 00:03:55.819 --> 00:03:58.960 heavy lifting complex reasoning, writing nuanced 00:03:58.960 --> 00:04:02.139 text. But for the simple stuff like basic classification 00:04:02.139 --> 00:04:05.199 or sorting, you program it to use the faster, 00:04:05.379 --> 00:04:08.379 much cheaper models like maybe GPT 4 .1 mini 00:04:08.379 --> 00:04:10.879 or something similar. You trade a tiny bit of 00:04:10.879 --> 00:04:13.400 setup complexity for potentially huge savings 00:04:13.400 --> 00:04:17.259 on running costs. So the agent essentially knows 00:04:17.259 --> 00:04:21.019 whether to use the cheap fast AI or the more 00:04:21.019 --> 00:04:23.399 powerful one based on the task gets handed from 00:04:23.399 --> 00:04:26.620 that branching logic. dictates the model choice, 00:04:26.939 --> 00:04:29.000 optimizing for cost. OK, let's simplify for a 00:04:29.000 --> 00:04:30.319 second and actually build something or at least 00:04:30.319 --> 00:04:32.839 walk through it. That multipath AI learning helper 00:04:32.839 --> 00:04:34.600 agent you mentioned, the one that sorts user 00:04:34.600 --> 00:04:36.939 questions. Right. So step one, start trigger. 00:04:37.180 --> 00:04:39.279 We set that to text input because, well, the 00:04:39.279 --> 00:04:41.240 user's typing a question. Simple enough. Step 00:04:41.240 --> 00:04:43.699 two is the sorting agent, the classifier. Its 00:04:43.699 --> 00:04:45.920 job is to read that question and slot it to one 00:04:45.920 --> 00:04:49.120 of maybe four buckets, AI news, AI tool info, 00:04:49.439 --> 00:04:52.279 AI basics, or AI business ideas. And this brings 00:04:52.279 --> 00:04:54.990 us to a really crucial technical point. Structured 00:04:54.990 --> 00:04:57.529 output. You have to instruct this agent very 00:04:57.529 --> 00:05:00.310 specifically to output its decision in JSON format. 00:05:00.509 --> 00:05:03.430 It's kind of mandatory for reliable flow. Why 00:05:03.430 --> 00:05:06.550 JSON though? Why is that specific format so important? 00:05:06.850 --> 00:05:09.649 Can't the agent just, you know, output text saying 00:05:09.649 --> 00:05:13.560 category AI news? Because the next block in the 00:05:13.560 --> 00:05:15.399 chain, the conditional logic block, it needs 00:05:15.399 --> 00:05:18.519 something unambiguous. It can't easily or reliably 00:05:18.519 --> 00:05:21.920 parse natural language, which can be fuzzy. JSON 00:05:21.920 --> 00:05:25.199 -like category A -news forces a clean machine 00:05:25.199 --> 00:05:27.980 -readable structure. It removes the AI's creativity 00:05:27.980 --> 00:05:30.420 for that specific step, making the output totally 00:05:30.420 --> 00:05:32.560 predictable, which is essential for the workflow 00:05:32.560 --> 00:05:34.660 stability. OK, that makes sense. It guarantees 00:05:34.660 --> 00:05:37.120 the next step gets exactly what it expects, which 00:05:37.120 --> 00:05:39.259 leads right into step three, conditional logic. 00:05:39.399 --> 00:05:41.500 This is the traffic cop, basically, taking that 00:05:41.500 --> 00:05:44.579 precise JSON output and sending the request down 00:05:44.579 --> 00:05:47.439 the right path. Exactly. And each path leads 00:05:47.439 --> 00:05:50.040 to its own dedicated, super specialized agent 00:05:50.040 --> 00:05:53.160 with really specific instructions, like the AI 00:05:53.160 --> 00:05:55.740 news agent. It must use web search. It must find 00:05:55.740 --> 00:05:57.759 the five most important stories from the last 00:05:57.759 --> 00:05:59.879 week. And it must give you the headline, a short 00:05:59.879 --> 00:06:03.699 summary, and the source link. Very precise. Constraints 00:06:03.699 --> 00:06:06.079 are definitely key there. And the AI tool agent, 00:06:06.540 --> 00:06:08.819 it has to use web search to explain the tool, 00:06:08.839 --> 00:06:11.860 but also give three real concrete examples of 00:06:11.860 --> 00:06:13.660 how you'd use it to actually create something, 00:06:13.959 --> 00:06:16.279 not just theory. Then you've got the AI basics 00:06:16.279 --> 00:06:19.639 agent. The instruction here is cool. Explain 00:06:19.639 --> 00:06:22.139 things like a patient eighth grade teacher. Use 00:06:22.139 --> 00:06:25.199 everyday words, zero jargon, and this one doesn't 00:06:25.199 --> 00:06:27.560 even need web search access. And finally, the 00:06:27.560 --> 00:06:30.279 AI business ideas agent. Find three successful 00:06:30.279 --> 00:06:33.300 business ideas using AI right now. Explain how 00:06:33.300 --> 00:06:35.680 AI is used and give a real company example for 00:06:35.680 --> 00:06:38.019 each. Whoa. I mean, just imagine the precision 00:06:38.019 --> 00:06:40.459 needed for that one conditional block to reliably 00:06:40.459 --> 00:06:42.439 route, I don't know, maybe billions of queries 00:06:42.439 --> 00:06:44.740 over time purely based on those instructions 00:06:44.740 --> 00:06:47.540 feeding it clean JSON. That's real operational 00:06:47.540 --> 00:06:49.860 scale right there. So going back to the instructions 00:06:49.860 --> 00:06:52.500 for those final agents, why do they need to be 00:06:52.500 --> 00:06:54.860 so specific about the output format and even 00:06:54.860 --> 00:06:56.720 the style, like the eighth grade teacher part? 00:06:56.959 --> 00:06:59.470 Yeah, that's specificity. stops you from getting 00:06:59.470 --> 00:07:03.209 vague, unhelpful answers. It ensures consistency, 00:07:03.430 --> 00:07:05.389 which is really important for the person using 00:07:05.389 --> 00:07:08.550 it, you know, the end -user experience. Mid -roll 00:07:08.550 --> 00:07:10.470 sponsor read provided separately. Okay, let's 00:07:10.470 --> 00:07:12.629 power this up. Building the internal logic is 00:07:12.629 --> 00:07:15.269 one thing, but making these agents truly useful 00:07:15.269 --> 00:07:17.709 often means connecting them to the outside world. 00:07:17.879 --> 00:07:19.959 Right? Absolutely. And that's where MCP comes 00:07:19.959 --> 00:07:22.879 in, Model Context Protocol. Think of it simply 00:07:22.879 --> 00:07:25.199 as the secure handshake that lets these agents 00:07:25.199 --> 00:07:28.480 talk to external apps and services, stuff outside 00:07:28.480 --> 00:07:31.139 the OpenAI environment. So the agent is smart, 00:07:31.720 --> 00:07:33.319 but it's kind of stuck in its box until you use 00:07:33.319 --> 00:07:36.680 MCP. How does that work securely? MCP is the 00:07:36.680 --> 00:07:39.220 bridge. It enables connections to things like 00:07:39.220 --> 00:07:41.879 Gmail so it can read or send emails, Google Calendar 00:07:41.879 --> 00:07:44.959 for scheduling, e -commerce platforms like Shopify, 00:07:45.579 --> 00:07:48.220 payment processors like Stripe, and maybe most 00:07:48.220 --> 00:07:50.920 powerfully, Zapier, which connects to literally 00:07:50.920 --> 00:07:54.339 thousands of other apps. Wow, okay. That unlocks 00:07:54.339 --> 00:07:56.680 some seriously advanced possibilities then. Totally, 00:07:56.720 --> 00:07:59.500 like that email replying agent example. It could 00:07:59.500 --> 00:08:02.160 read incoming emails, use its logic to sort them, 00:08:02.160 --> 00:08:04.860 maybe urgent, needs review, just a notification, 00:08:05.439 --> 00:08:07.920 then draft detailed replies based on rules you 00:08:07.920 --> 00:08:10.259 set, and either send them automatically or just 00:08:10.259 --> 00:08:12.959 save them as drafts for a human to quickly check. 00:08:13.040 --> 00:08:15.759 Okay, but hold on. If we're giving an automated 00:08:15.759 --> 00:08:19.019 agent access to read and write in my Gmail or 00:08:19.019 --> 00:08:23.620 see my Shopify orders, what about security? Governance. 00:08:23.759 --> 00:08:26.199 That seems like a big step. That's a super critical 00:08:26.199 --> 00:08:28.180 question. And the system is built with that in 00:08:28.180 --> 00:08:32.019 mind. It requires very specific scoped permissions 00:08:32.019 --> 00:08:33.820 for every connection you set up. You have to 00:08:33.820 --> 00:08:36.139 think about that governance layer up front, setting 00:08:36.139 --> 00:08:38.299 rules that stop the agent from doing things it 00:08:38.299 --> 00:08:39.720 shouldn't. Like you wouldn't want your customer 00:08:39.720 --> 00:08:42.120 return agent to suddenly be able to read HR emails, 00:08:42.159 --> 00:08:44.299 right? So you define those permissions carefully 00:08:44.299 --> 00:08:46.360 before you let it run wild. Let's circle back 00:08:46.360 --> 00:08:48.600 to instruction clarity because honestly that 00:08:48.600 --> 00:08:50.720 feels like the hardest part for me too. I still 00:08:50.720 --> 00:08:53.559 wrestle with prompt drift myself sometimes. Getting 00:08:53.559 --> 00:08:57.080 instructions concise but also completely unambiguous. 00:08:57.320 --> 00:08:59.559 It's way harder than it sounds. Oh, I agree. 00:08:59.740 --> 00:09:01.720 It's a common struggle that vulnerability is 00:09:01.720 --> 00:09:04.000 real. It's the difference between a bad instruction 00:09:04.000 --> 00:09:06.960 like help with emails, which is useless, versus 00:09:06.960 --> 00:09:10.159 a good one. Read incoming customer support emails, 00:09:10.480 --> 00:09:13.100 identify the core problem described, write a 00:09:13.100 --> 00:09:15.340 friendly and empathetic reply under 100 words, 00:09:15.519 --> 00:09:18.740 and then update the CRM record to resolved. That 00:09:18.740 --> 00:09:21.320 specificity is the design work. And you mentioned 00:09:21.320 --> 00:09:23.620 model choice earlier. That's critical, too, for 00:09:23.620 --> 00:09:25.620 performance but also cost, right? Absolutely 00:09:25.620 --> 00:09:28.440 crucial. Use your top -tier model, maybe GPT 00:09:28.440 --> 00:09:30.799 -5 when available, for the really tough reasoning 00:09:30.799 --> 00:09:33.720 or generating complex creative text. But for 00:09:33.720 --> 00:09:36.600 simple sorting, classification, data extraction, 00:09:37.340 --> 00:09:39.820 stick to the faster, much cheaper models like 00:09:39.820 --> 00:09:42.879 GPT -4 .1 Mini. If you build an agent that's 00:09:42.879 --> 00:09:45.299 inefficient, maybe it loops unnecessarily or 00:09:45.299 --> 00:09:47.200 defaults to the expensive model for everything, 00:09:47.580 --> 00:09:49.720 you'll see those costs spike really fast. So 00:09:49.720 --> 00:09:52.320 if things do go wrong, the agent breaks, gives 00:09:52.320 --> 00:09:55.960 weird answers, or it's just slow, what are the 00:09:55.960 --> 00:09:58.279 quick troubleshooting checks? First place I'd 00:09:58.279 --> 00:10:00.799 look is often the conditional logic, the if -then 00:10:00.799 --> 00:10:04.200 stuff. If it's giving wrong answers, reread those 00:10:04.200 --> 00:10:07.000 rules carefully. Maybe the intent classification 00:10:07.000 --> 00:10:09.340 is slightly off. If it's just not answering, 00:10:09.559 --> 00:10:11.519 check your block connections are solid and make 00:10:11.519 --> 00:10:14.580 sure it's actually published. If it's slow, try 00:10:14.580 --> 00:10:17.360 reducing how much it relies on web search or 00:10:17.360 --> 00:10:19.919 swap in a faster model for simpler steps. If 00:10:19.919 --> 00:10:22.159 someone's just starting out building their very 00:10:22.159 --> 00:10:24.519 first simple agent, what's the single most important 00:10:24.519 --> 00:10:26.399 thing they should do right after building it? 00:10:26.559 --> 00:10:29.419 Test it with weird inputs. Seriously. See how 00:10:29.419 --> 00:10:32.080 it handles unexpected questions, confusing language, 00:10:32.340 --> 00:10:35.080 or maybe missing information? Always anticipate 00:10:35.080 --> 00:10:37.399 how users might break it. Okay, let's bring this 00:10:37.399 --> 00:10:39.559 all together. The really big idea here seems 00:10:39.559 --> 00:10:43.230 to be this. Democratization. The power to build 00:10:43.230 --> 00:10:45.570 sophisticated tools isn't just for coders anymore. 00:10:46.129 --> 00:10:48.330 It's accessible to, you know, the business designer, 00:10:48.409 --> 00:10:51.389 the process owner. Exactly. And the practical 00:10:51.389 --> 00:10:55.070 uses are just huge. Personal research assistants, 00:10:55.669 --> 00:10:58.529 automated customer support flows, content strategy 00:10:58.529 --> 00:11:01.309 helpers, even sales assistants that could help 00:11:01.309 --> 00:11:04.409 qualify leads or schedule demos. And that final 00:11:04.409 --> 00:11:06.389 piece of advice from the material really resonates. 00:11:06.970 --> 00:11:09.779 Start simple. Pick one clear, well -defined job 00:11:09.779 --> 00:11:12.340 for your agent. Test it like crazy. Then and 00:11:12.340 --> 00:11:14.500 only then, start adding more features of complexity. 00:11:14.820 --> 00:11:16.399 Yeah, so the challenge to you, the listener, 00:11:16.720 --> 00:11:20.259 is maybe. Open up the builder, find that one 00:11:20.259 --> 00:11:22.480 repetitive, maybe kind of boring task you do 00:11:22.480 --> 00:11:25.720 often, and try automating just that. That's where 00:11:25.720 --> 00:11:27.580 you'll likely see the quickest win. Which leaves 00:11:27.580 --> 00:11:29.299 us with a final thought, something to maybe chew 00:11:29.299 --> 00:11:31.419 on. Perhaps the biggest value of this agent builder 00:11:31.419 --> 00:11:34.000 isn't just that it makes automation easier. but 00:11:34.000 --> 00:11:36.460 that it forces us, maybe for the first time for 00:11:36.460 --> 00:11:38.379 some processes, to really define and understand 00:11:38.379 --> 00:11:40.320 the tasks we should be automating in the first 00:11:40.320 --> 00:11:40.600 place.