WEBVTT 00:00:00.000 --> 00:00:02.120 You know that immediate mental hurdle when you 00:00:02.120 --> 00:00:04.639 start thinking about an AI agent project? It's 00:00:04.639 --> 00:00:06.980 almost always complexity. Your mind just leaps 00:00:06.980 --> 00:00:11.019 to... Perfect prompts, multi -agent setups, rock 00:00:11.019 --> 00:00:13.699 -solid security. Right. And these ideas, they're 00:00:13.699 --> 00:00:15.900 important down the line, sure, but they kill, 00:00:15.919 --> 00:00:18.280 what, 90 % of projects before they even get off 00:00:18.280 --> 00:00:20.660 the ground? It's total analysis paralysis. It 00:00:20.660 --> 00:00:22.140 feels like you should be thinking about that 00:00:22.140 --> 00:00:25.199 stuff like it's best practice. But the reality, 00:00:25.399 --> 00:00:28.160 the sort of liberating truth here, is that the 00:00:28.160 --> 00:00:30.399 90 % you actually need to build something that 00:00:30.399 --> 00:00:32.359 works, something you can ship. You can probably 00:00:32.359 --> 00:00:35.500 grasp that in, like, under an hour. So welcome 00:00:35.500 --> 00:00:38.969 to the deep dive. Our goal here is to guide you 00:00:38.969 --> 00:00:41.409 through what really matters with AI agents quickly 00:00:41.409 --> 00:00:44.070 without you getting lost in all the noise. Yeah, 00:00:44.109 --> 00:00:46.070 we want to cut right through that perfectionism 00:00:46.070 --> 00:00:49.210 trap. Isolate those practical ship it today basics. 00:00:49.549 --> 00:00:53.289 Exactly. So today we're going to define the four 00:00:53.289 --> 00:00:56.789 core parts of any agent. Lay out a simple three 00:00:56.789 --> 00:00:59.250 -step launch sequence and cover the essential 00:00:59.250 --> 00:01:02.570 lightweight ways to handle tools, security, and 00:01:02.570 --> 00:01:05.090 actually deploying this thing. Okay, let's dive 00:01:05.090 --> 00:01:07.310 in and try to make this feel a bit less daunting. 00:01:07.709 --> 00:01:11.590 So segment one is this perfectionism trap. I 00:01:11.590 --> 00:01:13.670 think we've all felt this. You get an idea for 00:01:13.670 --> 00:01:16.090 an agent and immediately you're picturing how, 00:01:16.269 --> 00:01:19.319 I don't know. Google or OpenAI would build it, 00:01:19.340 --> 00:01:21.719 this massive, intricate system. And that's the 00:01:21.719 --> 00:01:23.480 moment, right? That's usually where the idea 00:01:23.480 --> 00:01:26.420 just stalls. Yeah. Because it's this engineering 00:01:26.420 --> 00:01:28.340 exercise for a scale you don't even have yet. 00:01:28.480 --> 00:01:31.500 Yeah. But the pattern for success, what we actually 00:01:31.500 --> 00:01:33.840 see working, it's developers starting really 00:01:33.840 --> 00:01:36.480 simple. Shipping fast. Shipping fast. And then 00:01:36.480 --> 00:01:39.579 iterating based on what real users are telling 00:01:39.579 --> 00:01:42.420 them. not based on some abstract theory of perfection. 00:01:42.700 --> 00:01:46.140 So the core message is really resist that temptation. 00:01:46.400 --> 00:01:49.140 Focus on the practical 90%, just what you need 00:01:49.140 --> 00:01:51.760 to launch something basic that works. What do 00:01:51.760 --> 00:01:54.980 you think is the main psychological block there? 00:01:55.019 --> 00:01:57.840 What stops people shipping? It's overcomplicating 00:01:57.840 --> 00:02:00.299 that foundation. That's what kills projects before 00:02:00.299 --> 00:02:03.200 they start. Simplicity, especially early on, 00:02:03.299 --> 00:02:05.680 that's actually a feature, not a bug. Okay, so 00:02:05.680 --> 00:02:08.219 let's talk anatomy. Before touching any code, 00:02:08.419 --> 00:02:10.360 what are the absolute essential parts? You're 00:02:10.360 --> 00:02:12.759 saying there are just four. Just four core components, 00:02:12.900 --> 00:02:14.539 yeah. Yeah. Doesn't matter how fancy it gets 00:02:14.539 --> 00:02:16.759 later, these are the building blocks. First up, 00:02:16.879 --> 00:02:20.759 tools. Think of these as the agent's hands. Okay. 00:02:20.919 --> 00:02:23.180 Their functions, letting the agent, you know, 00:02:23.199 --> 00:02:25.680 interact with the outside world, search the web, 00:02:25.860 --> 00:02:28.099 query a database, send an email. Right. Without 00:02:28.099 --> 00:02:29.919 tools, it's just talking. It can't do anything. 00:02:30.039 --> 00:02:32.400 Exactly. It's just a chatbot otherwise. Number 00:02:32.400 --> 00:02:36.129 two, the large language model, the LLM. That's 00:02:36.129 --> 00:02:39.889 the brain. Your GPT -clawed Gemini. Right. The 00:02:39.889 --> 00:02:42.469 reasoning engine, it looks at the user's request, 00:02:42.710 --> 00:02:44.669 looks at its instructions, and decides, okay, 00:02:44.770 --> 00:02:47.490 which tool do I need for this? Then number three 00:02:47.490 --> 00:02:50.110 is the system prompt, the instruction manual. 00:02:50.430 --> 00:02:51.870 Yeah, that's a good way to put it. It's the high 00:02:51.870 --> 00:02:54.449 -level programming. Defines the agent's personality. 00:02:54.689 --> 00:02:58.689 Is it helpful? Snarky. Sets its goals, its rules 00:02:58.689 --> 00:03:00.610 of engagement. And the fourth piece is memory 00:03:00.610 --> 00:03:04.520 systems, the context. Crucial. This covers short 00:03:04.520 --> 00:03:06.699 -term memory, basically, the current conversation 00:03:06.699 --> 00:03:09.860 history, and potentially long -term memory, like 00:03:09.860 --> 00:03:12.360 remembering user preferences from last week or 00:03:12.360 --> 00:03:15.159 important facts it learned. And that's it. Everything 00:03:15.159 --> 00:03:17.139 else, deployment, monitoring, that's all just 00:03:17.139 --> 00:03:20.360 refinement. It's refinement, not the core foundation. 00:03:20.639 --> 00:03:24.180 The LLM is the brain, sure, but which component 00:03:24.180 --> 00:03:26.319 actually lets it do things in the real world? 00:03:26.639 --> 00:03:29.159 That has to be the tools. They're the hands letting 00:03:29.159 --> 00:03:32.020 it interact with APIs, databases, whatever it 00:03:32.020 --> 00:03:34.599 needs. Okay, so how do we actually launch this? 00:03:34.919 --> 00:03:38.879 The hello world for agents. It's just three steps. 00:03:39.139 --> 00:03:42.439 Step one. Pick an LLM for prototyping. Just pick 00:03:42.439 --> 00:03:44.539 something cheap and fast. Don't agonize over 00:03:44.539 --> 00:03:47.039 it. Cloud Haiku is great for this. Okay, cheap 00:03:47.039 --> 00:03:49.560 and fast. Step two. Write a basic, clear system 00:03:49.560 --> 00:03:51.759 prompt. Just the essentials. You'll refine it 00:03:51.759 --> 00:03:53.400 later once you see how people actually use it. 00:03:53.460 --> 00:03:55.960 And step three. Add one tool. for just one to 00:03:55.960 --> 00:03:58.340 start maybe a simple web search or even just 00:03:58.340 --> 00:04:00.280 a calculator and if you do those three things 00:04:00.280 --> 00:04:02.580 you have a working agent you can take an instruction 00:04:02.580 --> 00:04:06.699 reason about it and perform an action and honestly 00:04:06.699 --> 00:04:10.580 this can be like less than 50 lines of code it's 00:04:10.580 --> 00:04:13.250 really not complex to start Okay, you mentioned 00:04:13.250 --> 00:04:15.110 setting up the LLM connection. You talked about 00:04:15.110 --> 00:04:17.389 OpenRouter. Why is that important early on? Ah, 00:04:17.389 --> 00:04:19.750 yeah, good question. OpenRouter basically gives 00:04:19.750 --> 00:04:23.290 you one API key that works for almost all the 00:04:23.290 --> 00:04:28.550 major LLMs, GPT, Cloud Gemini, Mistral, you name 00:04:28.550 --> 00:04:31.230 it. So switching is easy. Super easy. You change 00:04:31.230 --> 00:04:33.269 like one line of text in your config file, and 00:04:33.269 --> 00:04:35.470 boom, you're using a different model. No vendor 00:04:35.470 --> 00:04:38.069 lock -in, easy testing. It's a smart move from 00:04:38.069 --> 00:04:40.620 day one. You know, I have to admit, I still sometimes 00:04:40.620 --> 00:04:43.199 wrestle with the temptation to just let the LLM 00:04:43.199 --> 00:04:45.800 handle things like math. Skip writing a calculator 00:04:45.800 --> 00:04:47.540 tool. Oh, totally. It feels like it should be 00:04:47.540 --> 00:04:49.759 smart enough, right? Yeah. That's a classic trap. 00:04:50.259 --> 00:04:52.540 LLMs are fundamentally token prediction machines. 00:04:52.899 --> 00:04:56.420 They predict the next likely word or token. That 00:04:56.420 --> 00:04:58.740 makes them surprisingly bad at precise arithmetic, 00:04:58.959 --> 00:05:01.399 like really bad sometimes. So adding a simple 00:05:01.399 --> 00:05:03.660 calculator tool isn't just a nice -to -have, 00:05:03.800 --> 00:05:06.740 it's often necessary. It compensates for a core 00:05:06.740 --> 00:05:10.680 limitation of the LLM itself. Why is adding a 00:05:10.680 --> 00:05:13.879 tool for something basic like arithmetic so crucial 00:05:13.879 --> 00:05:17.220 then? Because LLMs are just inherently weak at 00:05:17.220 --> 00:05:19.779 precise math. They need tools to make up for 00:05:19.779 --> 00:05:21.980 that fundamental gap. Okay, let's nail down some 00:05:21.980 --> 00:05:26.029 specifics. LLM choices. For prototyping, like 00:05:26.029 --> 00:05:28.870 I said, I really like Cloud Haiku 4 .5. It's 00:05:28.870 --> 00:05:31.509 super cheap, incredibly fast, great for just 00:05:31.509 --> 00:05:33.310 getting things working. And then when you think 00:05:33.310 --> 00:05:35.689 about maybe moving towards production. Cloud 00:05:35.689 --> 00:05:37.889 Sonnet 4 .5 seems to be hitting a sweet spot 00:05:37.889 --> 00:05:39.990 right now. Good balance of intelligence, speed, 00:05:40.129 --> 00:05:42.550 and cost. But again, the real point here is use 00:05:42.550 --> 00:05:44.709 something like OpenRouter. Right. So switching 00:05:44.709 --> 00:05:47.810 between Haiku, Sonnet, maybe even GPT -4 is just 00:05:47.810 --> 00:05:50.180 changing that one line. Don't get locked in because 00:05:50.180 --> 00:05:52.819 of hype. Test what works for your use case. Makes 00:05:52.819 --> 00:05:55.040 sense. And the system prompt, that instruction 00:05:55.040 --> 00:05:57.519 manual. You mentioned a template to avoid staring 00:05:57.519 --> 00:05:59.800 at a blank page. Yeah, a simple five -section 00:05:59.800 --> 00:06:01.420 template really helps structure your thinking. 00:06:01.579 --> 00:06:04.240 What are the core sections? First three are key 00:06:04.240 --> 00:06:08.220 to start. One, persona and goals, just two or 00:06:08.220 --> 00:06:11.519 three sentences. Two, tool instructions, how 00:06:11.519 --> 00:06:14.459 and when to use the tools you've given it. Three, 00:06:15.000 --> 00:06:17.639 output format tone. How detailed should it be? 00:06:17.720 --> 00:06:18.939 That sort of thing. And the other two sections. 00:06:19.040 --> 00:06:21.040 You said ignore them at first. Right. Section 00:06:21.040 --> 00:06:24.519 four is for examples really only needed for complex, 00:06:24.779 --> 00:06:27.720 multi -step tool workflows. And section five 00:06:27.720 --> 00:06:29.639 is miscellaneous instructions. This is like the 00:06:29.639 --> 00:06:32.579 fix it section. So why start with just the first 00:06:32.579 --> 00:06:35.139 three and ignore the others initially? Because 00:06:35.139 --> 00:06:37.079 you should only fill in that miscellaneous section 00:06:37.079 --> 00:06:39.740 based on observed behavior. Let your agent mess 00:06:39.740 --> 00:06:42.120 up in testing. If it tries scheduling meetings 00:06:42.120 --> 00:06:44.870 at 3 a .m., then you add a rule. Never schedule 00:06:44.870 --> 00:06:48.230 meetings before 9 a .m. Use real feedback. Don't 00:06:48.230 --> 00:06:50.550 guess about edge cases up front. Start simple. 00:06:51.269 --> 00:06:54.189 Okay. Tool strategy. You mentioned a limit. Yeah. 00:06:54.230 --> 00:06:55.910 A general rule of thumb is try to keep it under 00:06:55.910 --> 00:06:58.389 10 tools per agent. Once you go beyond that, 00:06:58.449 --> 00:07:01.149 the LLM tends to get confused. The performance 00:07:01.149 --> 00:07:03.290 drops. It starts picking the wrong tool. It just 00:07:03.290 --> 00:07:05.610 gets overwhelmed. Only 10. I've seen people build 00:07:05.610 --> 00:07:08.029 agents with like dozens of tiny little function 00:07:08.029 --> 00:07:10.629 calls. Why the limit? It's really about the LLM's 00:07:10.629 --> 00:07:13.540 cognitive load and the context window. Every 00:07:13.540 --> 00:07:15.420 tool you add, you have to describe it in the 00:07:15.420 --> 00:07:18.800 prompt. That eats up precious token space. And 00:07:18.800 --> 00:07:21.300 the more tools it has to choose from, the more 00:07:21.300 --> 00:07:23.899 time and tokens it spends just deciding which 00:07:23.899 --> 00:07:26.519 one to use. Often incorrectly if there are too 00:07:26.519 --> 00:07:29.160 many similar options. Keep it focused. And if 00:07:29.160 --> 00:07:31.500 you could only build out one core capability 00:07:31.500 --> 00:07:35.540 at first. RJ. No question. Retrieval Augmented 00:07:35.540 --> 00:07:37.959 Generation. That's giving the agent access to 00:07:37.959 --> 00:07:40.120 search private documents, right? Internal knowledge 00:07:40.120 --> 00:07:42.399 basis. Exactly. And the data we're seeing suggests 00:07:42.399 --> 00:07:45.180 something like 80 % or more of the real business 00:07:45.180 --> 00:07:47.680 value from agents comes from this capability. 00:07:48.259 --> 00:07:50.279 Letting it use your company's internal data, 00:07:50.399 --> 00:07:52.939 customer history, technical docs. Master RRA 00:07:52.939 --> 00:07:55.680 first. Master RAG. And you unlock the most valuable 00:07:55.680 --> 00:07:58.139 use cases right out of the gate. Wow. Yeah, I 00:07:58.139 --> 00:08:00.259 can see that. Imagine just asking it to pull 00:08:00.259 --> 00:08:03.000 key risks from the last thousand customer contracts 00:08:03.000 --> 00:08:05.860 instantly. Yeah. Yeah, that's huge for a business. 00:08:06.019 --> 00:08:08.100 Totally transformative. Now, security. Don't 00:08:08.100 --> 00:08:10.680 panic. Just basic hygiene. First rule. Never, 00:08:10.759 --> 00:08:14.199 ever hard code API keys in your code. Use environment 00:08:14.199 --> 00:08:16.339 variables. Standard practice. Standard practice. 00:08:16.819 --> 00:08:20.180 Then look at Guardrails AI. It's an open source 00:08:20.180 --> 00:08:22.560 Python tool. It lets you basically wrap your 00:08:22.560 --> 00:08:24.899 agent. Wrap it? What does that do? It gives you 00:08:24.899 --> 00:08:27.490 input protection. It can block things like prompt 00:08:27.490 --> 00:08:30.889 injection attacks or filter out PII, you know, 00:08:30.930 --> 00:08:33.570 personally identifiable information like names, 00:08:33.710 --> 00:08:36.529 addresses. Okay, so it cleans the input. And 00:08:36.529 --> 00:08:38.419 it filters the output, too. It can check for 00:08:38.419 --> 00:08:40.799 factual consistency, make sure the agent isn't 00:08:40.799 --> 00:08:43.080 leaking sensitive data. It's like a basic safety 00:08:43.080 --> 00:08:45.740 net, lets you ship things internally with a lot 00:08:45.740 --> 00:08:48.019 more confidence. And you also mentioned SNCC 00:08:48.019 --> 00:08:51.460 for vulnerabilities. Yeah, there are prepackaged 00:08:51.460 --> 00:08:54.259 tool collections, those called MCP servers, that 00:08:54.259 --> 00:08:56.340 bundle security checks. Using something like 00:08:56.340 --> 00:08:58.379 SNCCs can automate vulnerability detection during 00:08:58.379 --> 00:09:00.519 development. Just good practice. Yeah. So if 00:09:00.519 --> 00:09:02.720 you're focusing on just one tool capability for 00:09:02.720 --> 00:09:05.139 that first agent, what's the priority? It's got 00:09:05.139 --> 00:09:07.779 to be our key. Because that ability to tap into 00:09:07.779 --> 00:09:10.240 private company data, that's where you'll find 00:09:10.240 --> 00:09:12.620 the most immediate business value. Okay, last 00:09:12.620 --> 00:09:15.539 lap. Optimization and actually shipping this 00:09:15.539 --> 00:09:18.679 thing. Cost is a big one. Context window costs. 00:09:18.980 --> 00:09:22.220 Right. You pay per token, both input and output. 00:09:22.740 --> 00:09:25.320 So keep your system prompts concise. Keep your 00:09:25.320 --> 00:09:27.779 tool descriptions tight. They get sent with every 00:09:27.779 --> 00:09:30.580 single call to the LLM. Don't be verbose there. 00:09:30.740 --> 00:09:32.710 And for memory. For the conversation history. 00:09:32.990 --> 00:09:35.149 You absolutely need that sliding window memory 00:09:35.149 --> 00:09:37.029 trick we talked about. Don't send the entire 00:09:37.029 --> 00:09:39.830 chat history back every time, especially if it's 00:09:39.830 --> 00:09:42.710 a long conversation. Just the last, say, 10 or 00:09:42.710 --> 00:09:46.230 20 messages. Exactly. Use that simple list slicing 00:09:46.230 --> 00:09:49.830 like conversation 10 in Python. It dramatically 00:09:49.830 --> 00:09:52.809 cuts down your token costs. What about things 00:09:52.809 --> 00:09:56.019 the agent needs to remember long term? like user 00:09:56.019 --> 00:09:59.159 preferences or facts it learned weeks ago. For 00:09:59.159 --> 00:10:01.440 that, you'd look at something like MEM0. It's 00:10:01.440 --> 00:10:03.639 an open source tool specifically for persistent 00:10:03.639 --> 00:10:07.120 memory. It uses AG principles itself to store 00:10:07.120 --> 00:10:09.600 facts sufficiently and only retrieve the relevant 00:10:09.600 --> 00:10:11.779 ones for the current query. So it doesn't stuff 00:10:11.779 --> 00:10:14.259 the main context window with old info. Right. 00:10:14.500 --> 00:10:17.320 Avoids those unnecessary token costs for stuff 00:10:17.320 --> 00:10:19.580 that's not immediately needed. Then there's observability, 00:10:19.700 --> 00:10:22.029 seeing what's going on inside. Yeah, you really 00:10:22.029 --> 00:10:24.350 need this for debugging and just understanding 00:10:24.350 --> 00:10:27.309 costs. LangFuse is a great, easy to integrate 00:10:27.309 --> 00:10:29.570 option. Gives you a dashboard. What does it track? 00:10:29.730 --> 00:10:31.850 Tracks the whole execution flow step by step. 00:10:32.070 --> 00:10:34.870 Shows you token usage per step, latency, which 00:10:34.870 --> 00:10:37.830 system prompt was used. Yeah. Invaluable when 00:10:37.830 --> 00:10:39.350 your agent does something weird and you need 00:10:39.350 --> 00:10:41.990 to figure out why. And finally, deployment, getting 00:10:41.990 --> 00:10:44.289 it running. Think Docker native right from the 00:10:44.289 --> 00:10:46.789 start. Yeah. Build your agent inside a Docker 00:10:46.789 --> 00:10:49.730 container. Makes it super portable. Is it heavy? 00:10:50.240 --> 00:10:52.820 Does it need big servers? That's the surprising 00:10:52.820 --> 00:10:55.379 part. AI agents are usually really lightweight. 00:10:55.799 --> 00:10:58.320 All the heavy computation, the LLM inference 00:10:58.320 --> 00:11:01.139 that happens on open AIs or anthropic servers. 00:11:01.500 --> 00:11:04.220 Oh, okay. Your code is mostly just managing API 00:11:04.220 --> 00:11:07.100 calls and maybe running a simple tool. So it 00:11:07.100 --> 00:11:09.039 can run on a small, cheap server. A Docker container, 00:11:09.259 --> 00:11:11.460 maybe with a basic web front end like Streamlit 00:11:11.460 --> 00:11:14.100 if it's chatbot. Or just a serverless function 00:11:14.100 --> 00:11:16.279 like EWS Lambda if it runs in the background. 00:11:16.460 --> 00:11:19.710 Keep it simple. So token costs, they add up fast. 00:11:20.090 --> 00:11:22.070 What's the absolute simplest, most effective 00:11:22.070 --> 00:11:25.169 way to manage memory for a long chat? Use that 00:11:25.169 --> 00:11:28.269 sliding window. Send only the last 10 or 20 messages. 00:11:28.590 --> 00:11:31.289 That saves huge amounts on token costs compared 00:11:31.289 --> 00:11:33.909 to sending the full history every time. Okay, 00:11:33.970 --> 00:11:36.370 let's recap the big idea here. The 90 -10 rule. 00:11:36.809 --> 00:11:39.570 What's the 90 % that really matters? The stuff 00:11:39.570 --> 00:11:41.960 you should do now. All right. Bic and LLM Haiku 00:11:41.960 --> 00:11:44.559 is great for testing. Write a basic system prompt, 00:11:44.679 --> 00:11:47.120 just those first three template sections. Add 00:11:47.120 --> 00:11:50.200 one to three tools. Really focus on ARG for accessing 00:11:50.200 --> 00:11:52.279 internal data. That's usually the highest value 00:11:52.279 --> 00:11:55.080 starting point. Add basic security with guardrails, 00:11:55.139 --> 00:11:58.600 AI input, and output protection. Add simple observability, 00:11:58.759 --> 00:12:01.080 something like LangFuse. Control that short -term 00:12:01.080 --> 00:12:03.820 memory cost using the sliding window trick. And 00:12:03.820 --> 00:12:05.860 build it all with Docker in mind from day one. 00:12:06.059 --> 00:12:08.740 Easy deployment later. And critically, what's 00:12:08.740 --> 00:12:11.200 the 10 % you should actively ignore at the start? 00:12:11.480 --> 00:12:14.620 Ignore the complex multi -agent systems. Ignore 00:12:14.620 --> 00:12:16.980 crafting 5 ,000 -word perfect system prompts. 00:12:17.399 --> 00:12:19.460 Ignore Kubernetes orchestration for your first 00:12:19.460 --> 00:12:21.919 simple agent. Ignore building massive custom 00:12:21.919 --> 00:12:24.480 evaluation test suites before you even have user 00:12:24.480 --> 00:12:27.200 feedback. Right. The agents that actually ship. 00:12:27.399 --> 00:12:30.120 They shipped because the builders resisted that 00:12:30.120 --> 00:12:32.740 urge to over -engineer everything up front. Exactly. 00:12:32.779 --> 00:12:35.000 Focus on the foundations, get it out there, get 00:12:35.000 --> 00:12:38.480 real feedback, then iterate. So the call to action 00:12:38.480 --> 00:12:41.100 is pretty clear. Find that 50 -line code example 00:12:41.100 --> 00:12:43.480 mentioned in the source material, or just start 00:12:43.480 --> 00:12:45.639 fresh with those core components. Build something 00:12:45.639 --> 00:12:48.220 simple, like today. Yeah, don't wait for it to 00:12:48.220 --> 00:12:50.460 be perfect. You don't fail by shipping something 00:12:50.460 --> 00:12:53.240 simple or slightly flawed. You fail by letting 00:12:53.240 --> 00:12:55.539 perfectionism stop you from learning. And you 00:12:55.539 --> 00:12:58.149 only really learn from real -world usage. So 00:12:58.149 --> 00:13:00.309 a final thought to leave you with. What simple 00:13:00.309 --> 00:13:02.909 problem, maybe even an annoying little task you 00:13:02.909 --> 00:13:05.289 do every day, could your first basic agent solve 00:13:05.289 --> 00:13:08.389 right now? Outro music fades in.