WEBVTT 00:00:00.000 --> 00:00:03.220 Imagine feeding a single book draft into an AI. 00:00:04.299 --> 00:00:06.580 Instantly, you get five completely different 00:00:06.580 --> 00:00:09.460 reviews back. One review is from Linda. She's 00:00:09.460 --> 00:00:12.240 a retired teacher, brand new to AI. Another is 00:00:12.240 --> 00:00:14.859 from David, a corporate executive. Right. It's 00:00:14.859 --> 00:00:18.179 the exact same text draft, but you get completely 00:00:18.179 --> 00:00:22.660 distinct parallel perspectives. It feels like 00:00:22.660 --> 00:00:25.559 actual magic, but it's really just smart engineering. 00:00:25.780 --> 00:00:27.579 It really is. It's a fascinating shift in how 00:00:27.579 --> 00:00:30.170 we interact. Okay, let's unpack this. Welcome 00:00:30.170 --> 00:00:32.770 to our deep dive today. We're exploring the hidden 00:00:32.770 --> 00:00:35.390 power of Claude Code's sub -agents. We're going 00:00:35.390 --> 00:00:36.950 to look at what they actually are. We'll see 00:00:36.950 --> 00:00:39.509 why they save massive time and money. Then we'll 00:00:39.509 --> 00:00:42.090 help you build a custom plan roaster. We'll also 00:00:42.090 --> 00:00:44.929 set up vital safety guardrails you need. And 00:00:44.929 --> 00:00:46.990 finally, we'll scale up to dynamic workflows. 00:00:47.450 --> 00:00:50.130 I'm joined by our expert guide today. Let's dive 00:00:50.130 --> 00:00:53.130 right in. Hey there. I'm incredibly excited to 00:00:53.130 --> 00:00:55.109 get into this topic. That opening scenario with 00:00:55.109 --> 00:00:56.929 the book reviews sounds incredible, but I want 00:00:56.929 --> 00:00:59.270 to understand the mechanics here first. How does 00:00:59.270 --> 00:01:01.030 that parallel work actually happen behind the 00:01:01.030 --> 00:01:03.189 scenes? Well, it's a beautifully simple system 00:01:03.189 --> 00:01:05.370 of delegation. You can think of the main chat 00:01:05.370 --> 00:01:07.750 as your boss. The sub -agents are the workers. 00:01:07.989 --> 00:01:10.730 You only ever talk directly to the boss. Sub 00:01:10.730 --> 00:01:12.870 -agents don't talk to you directly either. They 00:01:12.870 --> 00:01:14.530 don't even talk to each other. They just do their 00:01:14.530 --> 00:01:17.000 job and report back. So it's kind of like a bustling 00:01:17.000 --> 00:01:19.579 restaurant kitchen. The main session is the head 00:01:19.579 --> 00:01:22.040 chef talking to customers. The sub -agents are 00:01:22.040 --> 00:01:24.420 prep cooks chopping onions in the back. Yeah, 00:01:24.480 --> 00:01:26.900 that's a perfect analogy. And setting up your 00:01:26.900 --> 00:01:29.560 kitchen this way gives four massive benefits. 00:01:29.859 --> 00:01:33.159 The very first one is keeping a pristine context 00:01:33.159 --> 00:01:37.040 window. Long AI chats get polluted incredibly 00:01:37.040 --> 00:01:40.489 fast. Yeah, because you make the AI read up on 00:01:40.489 --> 00:01:43.090 stuff. Like reading about Firefly's AI, just 00:01:43.090 --> 00:01:45.810 for a quick summary, that junk stays in its memory 00:01:45.810 --> 00:01:48.409 forever. Nobody wants that useless noise clogging 00:01:48.409 --> 00:01:50.909 up the conversation. Exactly. Every single token 00:01:50.909 --> 00:01:53.069 matters. The second benefit is a drastically 00:01:53.069 --> 00:01:56.150 lower overall cost. You use the really smart 00:01:56.150 --> 00:01:59.349 Opus model for planning, but you spin up cheaper 00:01:59.349 --> 00:02:02.290 Haiku models for heavy reading. That makes perfect 00:02:02.290 --> 00:02:04.109 financial sense right there. You don't pay a 00:02:04.109 --> 00:02:07.280 head chef to peel potatoes. Spot on. The third 00:02:07.280 --> 00:02:10.060 benefit is pure parallel work. You can review 00:02:10.060 --> 00:02:13.020 15 book chapters at the exact same time. You 00:02:13.020 --> 00:02:14.759 could research five different competitors all 00:02:14.759 --> 00:02:18.219 at once. That speed is just incredible. I know 00:02:18.219 --> 00:02:21.039 trying 15 chapters normally makes the AI hallucinate, 00:02:21.159 --> 00:02:24.780 so parallel work clearly fixes that issue. But 00:02:24.780 --> 00:02:27.159 what about the actual quality of the feedback? 00:02:27.500 --> 00:02:30.699 That brings us right to the fourth huge benefit. 00:02:30.900 --> 00:02:34.770 You get a truly fresh review every time. Subagents 00:02:34.770 --> 00:02:36.949 always start with completely blank memories. 00:02:37.110 --> 00:02:39.729 They give you honest and unbiased feedback. They 00:02:39.729 --> 00:02:41.550 aren't trying to be agreeable. Is context pollution 00:02:41.550 --> 00:02:44.030 really that big of a deal in normal workflows? 00:02:44.289 --> 00:02:47.090 Yes. It's a massive problem. Forcing one model 00:02:47.090 --> 00:02:49.669 to hold every file degrades its focus. It loses 00:02:49.669 --> 00:02:52.270 its logic over time. So crowded memory makes 00:02:52.270 --> 00:02:55.270 the AI lose focus and logic. Exactly. The kitchen 00:02:55.270 --> 00:02:57.129 analogy makes total sense to me. I definitely 00:02:57.129 --> 00:02:59.469 want those prep cooks saving me time. But if 00:02:59.469 --> 00:03:01.819 I'm the head chef here. How do we actually hire 00:03:01.819 --> 00:03:03.819 them? How are they given their specific instructions? 00:03:04.159 --> 00:03:06.539 It's surprisingly simple under the hood. A custom 00:03:06.539 --> 00:03:08.539 subagent is literally just a single markdown 00:03:08.539 --> 00:03:11.539 file. You usually store it in a hidden .clodagents 00:03:11.539 --> 00:03:13.819 folder. Wait, hold on. Just a standard text file? 00:03:14.000 --> 00:03:16.180 I assumed we were talking about complex Python 00:03:16.180 --> 00:03:18.639 scripts here. Nope. Just a standard markdown 00:03:18.639 --> 00:03:22.139 file. It has two completely distinct parts. The 00:03:22.139 --> 00:03:25.099 top part is the YAML front matter. Let's pause 00:03:25.099 --> 00:03:27.800 for a second. Define YAML front matter for us. 00:03:27.979 --> 00:03:30.969 A small... settings block at the top of a file. 00:03:31.069 --> 00:03:33.830 Got it. It holds the name and the trigger description. 00:03:34.050 --> 00:03:36.610 It also holds the model you want to use. Just 00:03:36.610 --> 00:03:39.310 basic configuration then. So what is the second 00:03:39.310 --> 00:03:41.150 part of the file? The second part is the main 00:03:41.150 --> 00:03:44.150 body. That's the actual workflow. It tells the 00:03:44.150 --> 00:03:46.770 agent exactly how to think. It lists the exact 00:03:46.770 --> 00:03:49.729 steps and the required output format. I read 00:03:49.729 --> 00:03:51.830 that Claude uses progressive disclosure here. 00:03:51.990 --> 00:03:54.909 Why is that specific feature so important? Think 00:03:54.909 --> 00:03:58.000 about it like RAM on your computer. If Claude 00:03:58.000 --> 00:04:00.340 loaded massive instruction sets every time you 00:04:00.340 --> 00:04:02.699 said hello, it would just grind to a total halt. 00:04:03.379 --> 00:04:05.520 Progressive disclosure means it only reads that 00:04:05.520 --> 00:04:08.300 tiny trigger description first. Ah, I see. So 00:04:08.300 --> 00:04:10.319 if the trigger matches, it loads the full body. 00:04:10.479 --> 00:04:12.840 That saves a massive amount of processing power. 00:04:13.180 --> 00:04:15.900 It's exactly right. Let's look at a concrete 00:04:15.900 --> 00:04:18.680 example to make this real. We call it the plan 00:04:18.680 --> 00:04:21.420 roaster. I really love that name. It's a phenomenal 00:04:21.420 --> 00:04:24.730 tool. The trigger description is kept remarkably 00:04:24.730 --> 00:04:28.370 simple. It just says to use this agent to critique 00:04:28.370 --> 00:04:32.170 a plan. I have a confession to make here. I still 00:04:32.170 --> 00:04:34.910 wrestle with prompt drift and misfires myself. 00:04:35.250 --> 00:04:38.269 Beat. It's frustrating when the AI just ignores 00:04:38.269 --> 00:04:41.069 your tools. Here's where it gets really interesting, 00:04:41.170 --> 00:04:43.730 though. People fall into a psychological trap 00:04:43.730 --> 00:04:46.110 when building these. They make those trigger 00:04:46.110 --> 00:04:48.350 descriptions way too long. They really do. It's 00:04:48.350 --> 00:04:51.579 a very classic mistake. A long description confuses 00:04:51.579 --> 00:04:53.939 the whole routing system. You must keep it remarkably 00:04:53.939 --> 00:04:56.199 short and punchy. Put the real heavy lifting 00:04:56.199 --> 00:04:58.680 down on the body instead. Right. The body is 00:04:58.680 --> 00:05:01.319 where you get highly specific. You tell the agent 00:05:01.319 --> 00:05:04.000 to find the absolute weakest points. Yeah, and 00:05:04.000 --> 00:05:06.319 you tell it to explain what could go wrong. You 00:05:06.319 --> 00:05:08.759 want it to be totally ruthless. What happens 00:05:08.759 --> 00:05:11.060 if the sub -agent just doesn't fire when I want 00:05:11.060 --> 00:05:13.800 it to? Well, that's called a misfire. It usually 00:05:13.800 --> 00:05:16.079 means your description is too vague. You fix 00:05:16.079 --> 00:05:18.410 it by trimming the description down. If it ignores 00:05:18.410 --> 00:05:20.550 you, make the trigger description shorter and 00:05:20.550 --> 00:05:23.050 sharper. Precisely. Okay, we know how to build 00:05:23.050 --> 00:05:25.990 one now. But once we write this text file down, 00:05:26.290 --> 00:05:28.790 where do we actually put it? We don't want these 00:05:28.790 --> 00:05:30.870 prep cooks getting lost. You need to understand 00:05:30.870 --> 00:05:34.829 skills versus subagents first. Skills run directly 00:05:34.829 --> 00:05:38.149 inside your main session context. Subagents run 00:05:38.149 --> 00:05:41.149 in a separate, fresh session entirely. But they 00:05:41.149 --> 00:05:43.230 aren't totally isolated from each other, right? 00:05:43.290 --> 00:05:44.790 They can actually work together. Absolutely. 00:05:44.850 --> 00:05:47.769 A subagent can actually invoke a skill. Right. 00:05:47.769 --> 00:05:50.529 Imagine spinning up five subagents at once. They 00:05:50.529 --> 00:05:52.430 all use a LinkedIn research skill simultaneously. 00:05:53.009 --> 00:05:55.569 That is an incredibly powerful combination. It's 00:05:55.569 --> 00:05:58.029 like giving your prep cooks power tools. It really 00:05:58.029 --> 00:06:00.329 is. Now, for storing these files, you have two 00:06:00.329 --> 00:06:02.850 main choices. You have project -level folders 00:06:02.850 --> 00:06:05.870 and global -level folders. Let's break down why 00:06:05.870 --> 00:06:08.230 you'd use one over the other. Project -level 00:06:08.230 --> 00:06:10.889 agents live inside your specific code repository. 00:06:11.750 --> 00:06:15.470 They're best for dedicated team workflows. Think 00:06:15.470 --> 00:06:17.889 about a strict security reviewer agent. That 00:06:17.889 --> 00:06:20.350 makes total sense. Because when you push the 00:06:20.350 --> 00:06:24.170 code to GitHub, your entire team gets that specific 00:06:24.170 --> 00:06:27.089 agent automatically. It stays directly with the 00:06:27.089 --> 00:06:30.310 project it belongs to. Exactly. Global agents 00:06:30.310 --> 00:06:32.889 live in your personal user directory instead. 00:06:33.029 --> 00:06:35.670 You use those across all your different projects. 00:06:36.009 --> 00:06:38.170 A personal writing reviewer is a great example. 00:06:38.490 --> 00:06:40.689 If I put an agent in the wrong folder, am I stuck? 00:06:41.029 --> 00:06:43.209 Not at all. Since they're literally just standard 00:06:43.209 --> 00:06:45.850 markdown files, you can just drag and drop them 00:06:45.850 --> 00:06:47.910 to a new folder later. They are just text files. 00:06:48.050 --> 00:06:50.029 You can move them anywhere, anytime. Exactly 00:06:50.029 --> 00:06:52.449 right. It's highly flexible. Midroll sponsor 00:06:52.449 --> 00:06:55.050 read placeholder. All right, let's get back to 00:06:55.050 --> 00:06:57.910 it. Okay, we have an army of cheap parallel workers 00:06:57.910 --> 00:07:00.050 now. We know how to build them and store them. 00:07:00.410 --> 00:07:03.470 But if we let them run wild, they could completely 00:07:03.470 --> 00:07:05.529 break our systems. We need to talk about security 00:07:05.529 --> 00:07:08.389 guardrails. Security is absolutely non -negotiable 00:07:08.389 --> 00:07:11.110 here. You have to adopt a zero -trust mindset. 00:07:11.740 --> 00:07:14.660 If an AI can touch something, assume it might. 00:07:14.819 --> 00:07:17.980 That is a slightly terrifying but deeply realistic 00:07:17.980 --> 00:07:21.040 rule. How do we actually enforce that zero trust 00:07:21.040 --> 00:07:23.860 policy? You must use strict tool restrictions. 00:07:24.279 --> 00:07:26.819 Always use read -only tools for your review agents. 00:07:27.339 --> 00:07:29.360 Our plan roaster doesn't need permission to write 00:07:29.360 --> 00:07:32.000 files. It only needs to read your plan and return 00:07:32.000 --> 00:07:34.939 feedback. So you only give edit access when strictly 00:07:34.939 --> 00:07:38.300 necessary. Never do it by default. Exactly. You 00:07:38.300 --> 00:07:40.620 also have to carefully limit your MCP access. 00:07:41.160 --> 00:07:43.620 Let's define MCP servers for the listener really 00:07:43.620 --> 00:07:46.079 quickly. Tools letting AI talk to your private 00:07:46.079 --> 00:07:48.560 databases and apps. Got it. So you don't give 00:07:48.560 --> 00:07:51.279 a writing critic database access. Please don't. 00:07:51.279 --> 00:07:54.079 That's a recipe for absolute disaster. Imagine 00:07:54.079 --> 00:07:56.399 a rogue subagent accidentally dropping your user 00:07:56.399 --> 00:07:59.220 tables. You also need to enforce max turns in 00:07:59.220 --> 00:08:01.379 your settings. What does the max turn setting 00:08:01.379 --> 00:08:04.600 actually do? It firmly caps how long an agent 00:08:04.600 --> 00:08:06.920 works. It prevents them from getting stuck in 00:08:06.920 --> 00:08:09.120 infinite loops. And it stops them from burning 00:08:09.120 --> 00:08:11.160 through your wallet. What about downloading community 00:08:11.160 --> 00:08:13.740 agents from public repositories? People really 00:08:13.740 --> 00:08:17.120 love sharing their custom setups online. Be incredibly 00:08:17.120 --> 00:08:19.240 careful with those. They're still just basic 00:08:19.240 --> 00:08:21.920 instruction files. They can easily contain malicious 00:08:21.920 --> 00:08:24.589 prompt injection or data leaks. You know what 00:08:24.589 --> 00:08:27.050 I think is a brilliant idea here? Building an 00:08:27.050 --> 00:08:30.329 AI agent specifically to act as a bouncer? Its 00:08:30.329 --> 00:08:33.029 entire job is just to inspect third -party agents. 00:08:33.269 --> 00:08:36.350 That's a fantastic safety practice. Build a read 00:08:36.350 --> 00:08:39.549 -only verifier subagent just for that exact purpose. 00:08:39.730 --> 00:08:42.029 It reviews the code before the boss ever sees 00:08:42.029 --> 00:08:44.769 it. Why isn't a prompt like, do not edit files 00:08:44.769 --> 00:08:48.350 enough to keep an agent safe? Because text instructions 00:08:48.350 --> 00:08:51.909 can be ignored or hallucinated away. Real permission 00:08:51.909 --> 00:08:54.649 limits physically block the AI from taking dangerous 00:08:54.649 --> 00:08:58.009 action. Never trust a polite request. Use hard 00:08:58.009 --> 00:09:00.169 permission blocks instead. That is the golden 00:09:00.169 --> 00:09:02.610 rule of agent security. So our system is safely 00:09:02.610 --> 00:09:05.590 locked down and isolated now. How far can we 00:09:05.590 --> 00:09:07.809 actually push this boss -worker relationship? 00:09:08.269 --> 00:09:10.750 What is the true ceiling here? We can push it 00:09:10.750 --> 00:09:13.370 remarkably far. Your main session can spin up 00:09:13.370 --> 00:09:15.669 massive numbers of sub -agents. We're not just 00:09:15.669 --> 00:09:17.409 talking about three or five agents. We're talking 00:09:17.409 --> 00:09:20.490 about serious dynamic scale. You can deploy 40. 00:09:20.889 --> 00:09:24.350 100 or even 200 agents. And they all run in perfect 00:09:24.350 --> 00:09:26.990 parallel simultaneously. I've seen developers 00:09:26.990 --> 00:09:29.330 trigger this by typing ultra code. It's kind 00:09:29.330 --> 00:09:32.009 of become a meme online. Yeah, that's a dramatic 00:09:32.009 --> 00:09:34.629 trigger phrase people use. The use cases for 00:09:34.629 --> 00:09:37.389 this scale are completely wild. You can review 00:09:37.389 --> 00:09:40.029 a massive legacy code base instantly. You can 00:09:40.029 --> 00:09:42.629 test dozens of experimental bug fixes all at 00:09:42.629 --> 00:09:45.149 once. Or review a massive thousand page book 00:09:45.149 --> 00:09:48.860 chapter by chapter. Two sec silence. Whoa, imagine 00:09:48.860 --> 00:09:51.600 scaling to 200 agents testing code -based fixes 00:09:51.600 --> 00:09:53.759 at once. That's a whole tech company inside your 00:09:53.759 --> 00:09:56.159 laptop. It really is. It fundamentally changes 00:09:56.159 --> 00:09:58.860 how fast you can iterate. But there's a very 00:09:58.860 --> 00:10:01.740 serious warning here. This burns through your 00:10:01.740 --> 00:10:04.720 session limits incredibly fast. It eats through 00:10:04.720 --> 00:10:07.259 your API tokens like crazy. Yeah, I can see how 00:10:07.259 --> 00:10:08.840 that gets expensive quickly. If you aren't paying 00:10:08.840 --> 00:10:12.389 attention, that's a massive bill. Exactly. Don't 00:10:12.389 --> 00:10:14.909 use dynamic workflows for small, trivial edits. 00:10:15.190 --> 00:10:17.850 Save this incredible power for the truly massive 00:10:17.850 --> 00:10:20.570 parallel tasks. When should I absolutely avoid 00:10:20.570 --> 00:10:23.450 using a subagent or dynamic workflow? Avoid them 00:10:23.450 --> 00:10:26.269 when the task is tiny, or if it depends heavily 00:10:26.269 --> 00:10:29.529 on previous chat history, or if it requires constant 00:10:29.529 --> 00:10:31.649 back -and -forth conversation with you. Keep 00:10:31.649 --> 00:10:34.230 small, highly conversational tasks in the main 00:10:34.230 --> 00:10:36.750 chat window. Exactly. Don't overcomplicate the 00:10:36.750 --> 00:10:39.259 simple stuff. So... Stepping back from all the 00:10:39.259 --> 00:10:41.480 technical execution, what does this all mean 00:10:41.480 --> 00:10:44.299 for us? If we connect this to the bigger picture, 00:10:44.460 --> 00:10:47.600 the core philosophy here is delegation. You have 00:10:47.600 --> 00:10:49.759 to change how you think. Ask yourself a simple 00:10:49.759 --> 00:10:53.220 question. Is this task going to dump a huge pile 00:10:53.220 --> 00:10:55.820 of text into my chat window? Stuff that I'll 00:10:55.820 --> 00:10:58.080 never actually read again. If the answer is yes, 00:10:58.220 --> 00:11:01.299 you delegate it immediately. Exactly. Keep the 00:11:01.299 --> 00:11:04.120 boss smart on Opus. Make the workers cheap on 00:11:04.120 --> 00:11:07.279 Haiku. Keep your context completely clean. And 00:11:07.279 --> 00:11:09.899 start small by building highly specific narrow 00:11:09.899 --> 00:11:12.639 specialists. We're moving into a really strange 00:11:12.639 --> 00:11:15.700 new era. You're no longer just an AI user typing 00:11:15.700 --> 00:11:18.779 prompts. You're an actual AI middle manager right 00:11:18.779 --> 00:11:20.740 now. Think about the implications of that for 00:11:20.740 --> 00:11:23.019 a second. What happens when your subagents eventually 00:11:23.019 --> 00:11:25.600 get the ability to hire their own subagents? 00:11:25.620 --> 00:11:27.500 At what point do we lose track of the bureaucracy? 00:11:27.940 --> 00:11:30.240 There's going to be this massive invisible bureaucracy 00:11:30.240 --> 00:11:33.399 happening inside our own computers. Beat. That 00:11:33.399 --> 00:11:35.779 is a mildly terrifying but utterly fascinating 00:11:35.779 --> 00:11:38.679 thought. It changes the whole definition of personal 00:11:38.679 --> 00:11:41.200 computing. It really does. I want to challenge 00:11:41.200 --> 00:11:43.600 you, the listener, go into your .cloud agents 00:11:43.600 --> 00:11:46.220 folder today. Just write one simple read -only 00:11:46.220 --> 00:11:48.320 subagent. See what happens when you delegate 00:11:48.320 --> 00:11:50.659 a tiny piece of your workflow. Thanks for joining 00:11:50.659 --> 00:11:53.019 us on this deep dive. Let those prep cooks get 00:11:53.019 --> 00:11:53.419 to work.