WEBVTT 00:00:00.000 --> 00:00:02.740 I spent the better part of last week just copying 00:00:02.740 --> 00:00:05.759 code snippets out of my IDE, pacing them into 00:00:05.759 --> 00:00:09.140 a chat window, waiting, and then carefully pacing 00:00:09.140 --> 00:00:11.259 the fix back in. It's so tedious, right? And 00:00:11.259 --> 00:00:14.060 it's fundamentally slow. Yeah. That realization, 00:00:14.359 --> 00:00:17.160 that friction, is the critical shift. We're still 00:00:17.160 --> 00:00:20.539 treating these incredible models like, you know, 00:00:20.699 --> 00:00:22.539 glorified chatbots. When they could be working 00:00:22.539 --> 00:00:25.440 directly in the project, autonomously. Exactly. 00:00:25.519 --> 00:00:27.859 We need to move beyond being the middleman. So 00:00:27.859 --> 00:00:31.510 this deep dive is about taking these technical 00:00:31.510 --> 00:00:34.009 guides we have, these blueprints, and forging 00:00:34.009 --> 00:00:36.929 them into a real strategy. The goal is to build 00:00:36.929 --> 00:00:40.579 an autonomous, multi -agent AI coding team. Our 00:00:40.579 --> 00:00:43.340 mission today for you is to extract the concrete 00:00:43.340 --> 00:00:47.079 tools, the specific GitHub workflows, and the 00:00:47.079 --> 00:00:48.960 crucial security measures you need to deploy 00:00:48.960 --> 00:00:51.460 a genuine 2 .0 .4 .7 development team. A team 00:00:51.460 --> 00:00:53.500 that lives inside your GitHub repository. And 00:00:53.500 --> 00:00:55.679 fixes issues while you're sleeping. OK, let's 00:00:55.679 --> 00:00:58.079 unpack this core idea first. Why bother with 00:00:58.079 --> 00:01:00.340 three different AI agents? Why not just use one 00:01:00.340 --> 00:01:03.079 massive model and prompt it to do everything? 00:01:03.219 --> 00:01:05.420 Isn't that just adding complexity? That's a great 00:01:05.420 --> 00:01:07.799 question, and it gets right to the heart of the 00:01:07.799 --> 00:01:11.760 architecture. The future here isn't one monolithic 00:01:11.760 --> 00:01:15.939 brain. It's a coordinated, specialized system. 00:01:16.340 --> 00:01:19.019 Think of it like a digital team. You want specific 00:01:19.019 --> 00:01:22.120 outputs for specific tasks, and different models 00:01:22.120 --> 00:01:25.099 give you different trade -offs. Speed, cost, 00:01:25.500 --> 00:01:28.390 and most importantly, control. Control being 00:01:28.390 --> 00:01:31.569 the kicker. Yes. Just like in a human team, you 00:01:31.569 --> 00:01:34.129 wouldn't hire your most creative architect to 00:01:34.129 --> 00:01:36.730 do a rigid, repetitive security audit. So we 00:01:36.730 --> 00:01:38.569 categorize them. So tell us what the lineup. 00:01:38.829 --> 00:01:40.930 Who's on the team? First up, we have the hybrid 00:01:40.930 --> 00:01:43.590 worker. We use Claude code for this. This is 00:01:43.590 --> 00:01:45.829 your best reasoning engine. It's great for complex 00:01:45.829 --> 00:01:48.849 logic, but, and this is key, human approval is 00:01:48.849 --> 00:01:50.650 required before it merges anything. So you're 00:01:50.650 --> 00:01:53.109 always in the loop. Safety first. Got it. Then 00:01:53.109 --> 00:01:55.569 you have the strict worker. For this, OpenAI 00:01:55.569 --> 00:01:57.750 Codex is ideal. This is where predictability 00:01:57.750 --> 00:02:01.049 is everything. Boring, repeatable tasks. Exactly. 00:02:01.230 --> 00:02:04.549 Writing unit tests, updating redomies, generating 00:02:04.549 --> 00:02:06.810 basic documentation. It has to follow a rigid, 00:02:07.069 --> 00:02:08.930 predictable process. And finally, the one that 00:02:08.930 --> 00:02:10.990 sounds the most fun. That's the fast worker. 00:02:11.590 --> 00:02:15.469 We use the cursor CLI for this. This is for complete 00:02:15.469 --> 00:02:20.150 autonomy. Full speed. It edits files, saves changes, 00:02:20.629 --> 00:02:23.030 and submits a pull request without you ever having 00:02:23.030 --> 00:02:26.060 to check in. Pure velocity. Pure velocity. So 00:02:26.060 --> 00:02:28.340 how does organizing them into these specialized 00:02:28.340 --> 00:02:32.020 roles actually improve efficiency over just using 00:02:32.020 --> 00:02:35.020 a single powerful model? Specialized roles give 00:02:35.020 --> 00:02:37.639 you precise control over the output. You can 00:02:37.639 --> 00:02:40.520 optimize for speed or rigidity as needed. Now, 00:02:40.539 --> 00:02:42.580 to make this system work, you need a foundation. 00:02:43.139 --> 00:02:45.949 If the agents are the workers, GitHub Actions 00:02:45.949 --> 00:02:49.150 is the robotic manager that tells them when to 00:02:49.150 --> 00:02:50.969 show up. Let's define that, Jorgan, quickly. 00:02:51.469 --> 00:02:53.569 GitHub Actions are basically a robotic butler 00:02:53.569 --> 00:02:55.610 for your code. Yeah, that's a great way to put 00:02:55.610 --> 00:02:57.729 it. They just follow a recipe when a specific 00:02:57.729 --> 00:02:59.370 event happens, like someone posting a comment. 00:02:59.629 --> 00:03:02.469 And that recipe needs three critical parts. First, 00:03:02.689 --> 00:03:05.069 the trigger. This is the magic word, maybe at 00:03:05.069 --> 00:03:07.870 CloudFix or at CursorFix, that you type as a 00:03:07.870 --> 00:03:09.889 comment on a GitHub issue. That wakes the bot 00:03:09.889 --> 00:03:12.069 up. Wakes the right bot up. Second, and this 00:03:12.069 --> 00:03:14.939 is crucial for security, is the runner. You're 00:03:14.939 --> 00:03:17.020 not running this on your laptop. GitHub gives 00:03:17.020 --> 00:03:19.360 you a temporary computer, a virtual machine, 00:03:19.759 --> 00:03:22.259 that spins up securely, does the work, and then 00:03:22.259 --> 00:03:24.879 just deletes itself. And the third part is the 00:03:24.879 --> 00:03:27.520 script itself, the workflow file. The recipe. 00:03:27.759 --> 00:03:30.280 It's a YAML text file that details all the steps. 00:03:30.659 --> 00:03:32.680 Read the comment, check who the user is, wake 00:03:32.680 --> 00:03:35.360 up the AI, send the code, save the changes. It's 00:03:35.360 --> 00:03:39.240 the conductor. Beyond security, What's the practical 00:03:39.240 --> 00:03:41.740 advantage of using a temporary GitHub runner 00:03:41.740 --> 00:03:44.340 instead of running these tasks locally? It prevents 00:03:44.340 --> 00:03:47.539 complex resource -intensive operations from slowing 00:03:47.539 --> 00:03:50.620 down your own computer. Okay, let's dive into 00:03:50.620 --> 00:03:53.860 method one, the safest approach, the hybrid model 00:03:53.860 --> 00:03:57.319 using Claude. The key here is that approval is 00:03:57.319 --> 00:03:59.500 always required. The human stays firmly in the 00:03:59.500 --> 00:04:01.139 loop. And that safety starts with permissions. 00:04:01.400 --> 00:04:03.740 It's non -negotiable. The first line of defense 00:04:03.740 --> 00:04:06.300 in that workflow file is a guest list, a defined 00:04:06.300 --> 00:04:08.639 list of authorized GitHub usernames. If anyone 00:04:08.639 --> 00:04:10.520 else tries to trigger the bot, the action just 00:04:10.520 --> 00:04:12.740 fails. Okay, so once you've secured who can use 00:04:12.740 --> 00:04:15.300 it, how do you instruct the bot without rewriting 00:04:15.300 --> 00:04:17.759 a huge prompt every single time? You definitely 00:04:17.759 --> 00:04:19.800 don't want to do that. You save a standard set 00:04:19.800 --> 00:04:22.420 of instructions in an instructions .md file. 00:04:22.939 --> 00:04:25.980 This tells Claude its role may be expert senior 00:04:25.980 --> 00:04:28.639 software engineer, and sets expectations for 00:04:28.639 --> 00:04:31.540 its tone. But just instructions for its personality 00:04:31.540 --> 00:04:33.779 isn't enough for code style, right? No, because 00:04:33.779 --> 00:04:36.000 the AI is only as good as the context you give 00:04:36.000 --> 00:04:38.920 it. And the critical trick here, and this is 00:04:38.920 --> 00:04:41.920 a constant battle, I still wrestle with prompt 00:04:41.920 --> 00:04:45.779 drift myself, is using an agents .md file in 00:04:45.779 --> 00:04:48.339 the repository. Tell me more about that specific 00:04:48.339 --> 00:04:50.620 file. That file is your project style guide, 00:04:50.639 --> 00:04:53.279 but for machines, it details your coding rules, 00:04:53.740 --> 00:04:56.279 always use TypeScript, indent with two spaces, 00:04:56.699 --> 00:04:59.600 all functions must be camel case. And every bot 00:04:59.600 --> 00:05:02.399 reads this first? Every single AI bot, no matter 00:05:02.399 --> 00:05:04.420 the model, is instructed to read that file first. 00:05:04.660 --> 00:05:07.160 This guarantees consistency. It stops the bots 00:05:07.160 --> 00:05:09.439 from contradicting each other's style. So what's 00:05:09.439 --> 00:05:12.290 the outcome of this hybrid approach? Claude reads 00:05:12.290 --> 00:05:14.430 the issue, reads the context, reads the style 00:05:14.430 --> 00:05:16.750 guide, creates a new Git branch with the fix, 00:05:17.129 --> 00:05:19.350 and then it comments back on the original issue 00:05:19.350 --> 00:05:21.610 with a clickable link to open the pull request 00:05:21.610 --> 00:05:24.449 for your final human review. So if the AI is 00:05:24.449 --> 00:05:27.949 so smart, why is using agents .md so crucial 00:05:27.949 --> 00:05:29.709 instead of just telling it the style once in 00:05:29.709 --> 00:05:33.110 a prompt? Contextual documentation ensures every 00:05:33.110 --> 00:05:35.509 bot consistently adheres to project standards. 00:05:35.829 --> 00:05:39.079 And that moves us to method two. strict and deterministic 00:05:39.079 --> 00:05:42.519 using OpenAI Codex. Here, the focus shifts to 00:05:42.519 --> 00:05:46.120 total uncompromising control. The AI only outputs 00:05:46.120 --> 00:05:48.540 text. The workflow controls everything else. 00:05:48.959 --> 00:05:51.079 I want to pause on that word, deterministic. 00:05:51.399 --> 00:05:54.339 That means predictable. Can AI -generated code 00:05:54.339 --> 00:05:56.339 really be deterministic, or does that just mean 00:05:56.339 --> 00:05:58.970 the process around the code is rigid? It's the 00:05:58.970 --> 00:06:01.290 latter. The process is absolutely rigid. The 00:06:01.290 --> 00:06:03.470 YAML file dictates everything. The branch name, 00:06:03.610 --> 00:06:05.509 the commit message where the file is saved, opening 00:06:05.509 --> 00:06:09.009 the PR. The AI's only job is to fill in the blanks 00:06:09.009 --> 00:06:11.639 with code. So you're just focusing its execution? 00:06:11.779 --> 00:06:13.680 Entirely. And the perfect use case for this is? 00:06:13.959 --> 00:06:16.560 Unit tests. Unit tests. It's boring, necessary 00:06:16.560 --> 00:06:18.680 work. A human doesn't want to spend two hours 00:06:18.680 --> 00:06:21.100 writing jest tests for every edge case. You set 00:06:21.100 --> 00:06:24.399 the AI's role to QA engineer, specify the framework, 00:06:24.839 --> 00:06:27.180 and demand exhaustive coverage. And what's the 00:06:27.180 --> 00:06:29.339 key to making sure the AI stays in that rigid 00:06:29.339 --> 00:06:31.879 lane? The prompt has to be explicit. It must 00:06:31.879 --> 00:06:35.490 include output. Only the code. Do not talk to 00:06:35.490 --> 00:06:38.769 me. Do not explain anything. Just the text. And 00:06:38.769 --> 00:06:41.290 the workflow script just grabs that raw output. 00:06:41.629 --> 00:06:44.189 Exactly. Captures that raw code and saves it 00:06:44.189 --> 00:06:46.529 directly into the right file like my -function 00:06:46.529 --> 00:06:49.850 .test .js before pushing it. No explanation needed. 00:06:50.189 --> 00:06:53.410 No conversation. So does this deterministic setup 00:06:53.410 --> 00:06:56.569 fundamentally limit the AI's ability to reason? 00:06:56.860 --> 00:06:59.500 Or does it just focus its execution? It focuses 00:06:59.500 --> 00:07:02.420 the execution, ensuring the output aligns perfectly 00:07:02.420 --> 00:07:04.720 with predictable file structures. Okay, now for 00:07:04.720 --> 00:07:07.379 maximum speed, this is method three, autonomous 00:07:07.379 --> 00:07:10.000 speed with the cursor CLI. This is where we run 00:07:10.000 --> 00:07:12.259 it in what's called headless mode. Right. Headless 00:07:12.259 --> 00:07:14.300 mode is just jargon for running a tool in the 00:07:14.300 --> 00:07:16.360 background without a user interface. It's perfect 00:07:16.360 --> 00:07:18.800 for an automated GitHub runner. We install the 00:07:18.800 --> 00:07:21.860 cursor CLI, and suddenly the AI has access to 00:07:21.860 --> 00:07:23.660 the command line. And because it has terminal 00:07:23.660 --> 00:07:26.480 access, you can skip a lot of the complex YAML 00:07:26.480 --> 00:07:28.680 scripting we needed for codecs. You just give 00:07:28.680 --> 00:07:31.319 it the do everything prompt. Precisely. You're 00:07:31.319 --> 00:07:33.899 basically telling the AI to manage the whole 00:07:33.899 --> 00:07:36.720 Git process itself. The prompt literally says 00:07:36.720 --> 00:07:39.560 things like, create a new branch called feature 00:07:39.560 --> 00:07:44.019 6, update the CSS file, verify the changes, commit 00:07:44.019 --> 00:07:47.560 them, and then use gapr create to open a pull 00:07:47.560 --> 00:07:50.379 request. So that's the power analogy here. You 00:07:50.379 --> 00:07:52.920 type in cursor fix, close your laptop, go make 00:07:52.920 --> 00:07:54.839 a coffee. And the PR is waiting for you when 00:07:54.839 --> 00:07:57.980 you get back. It's tireless, instantaneous development. 00:07:58.300 --> 00:08:01.829 Whoa. Just imagine scaling this across an entire 00:08:01.829 --> 00:08:04.949 company to handle thousands of repo updates instantly 00:08:04.949 --> 00:08:06.850 after a big security vulnerability is found. 00:08:07.230 --> 00:08:09.230 Humans can't match that speed. It's a different 00:08:09.230 --> 00:08:12.089 scale of maintenance. So since this is the fastest, 00:08:12.149 --> 00:08:14.509 most autonomous method, what practical steps 00:08:14.509 --> 00:08:17.209 should a manager take to audit its output effectively? 00:08:17.589 --> 00:08:19.829 Auditing is managed by integrating immediate 00:08:19.829 --> 00:08:22.470 automated security checks before the PR is even 00:08:22.470 --> 00:08:24.470 created. Before we move on, let's take a quick 00:08:24.470 --> 00:08:28.670 break. Welcome back to Deep Dive. We've talked 00:08:28.670 --> 00:08:30.310 about speed and autonomy, but when you give a 00:08:30.310 --> 00:08:32.309 machine that much power, security has to be the 00:08:32.309 --> 00:08:34.690 very next thought. Absolutely. If this AI team 00:08:34.690 --> 00:08:37.230 is working 24 -7, we need a security guard watching 00:08:37.230 --> 00:08:40.289 over them. And that guard is a tool like SonarCube. 00:08:40.590 --> 00:08:43.129 Exactly. It acts like an advanced spell checker 00:08:43.129 --> 00:08:45.669 for your code, finding bugs and critical security 00:08:45.669 --> 00:08:48.809 risks like SQL injection vulnerabilities. The 00:08:48.809 --> 00:08:51.750 key is integrating this scan before the code 00:08:51.750 --> 00:08:53.950 ever gets to a human reviewer. This is where 00:08:53.950 --> 00:08:56.549 the workflow gets really smart. The process is 00:08:56.549 --> 00:08:59.879 simple. AI writes the code, SonarKrub scans it, 00:09:00.179 --> 00:09:03.139 and if a security issue is found, SonarKrub immediately 00:09:03.139 --> 00:09:05.700 tells the AI that wrote it. to try again. It 00:09:05.700 --> 00:09:08.120 creates a self -correction loop. Right. The AI 00:09:08.120 --> 00:09:10.980 fixes the security issue on the spot, using that 00:09:10.980 --> 00:09:13.700 feedback, and only then is the pull request created. 00:09:14.059 --> 00:09:16.879 This means the code reaching a human is already 00:09:16.879 --> 00:09:20.059 pre -vetted and cleaner. It cuts down on so much 00:09:20.059 --> 00:09:22.480 wasted time. That dramatically improves the quality 00:09:22.480 --> 00:09:24.639 upstream. And my favorite concept from these 00:09:24.639 --> 00:09:26.960 guides has to be the triangle strategy. Making 00:09:26.960 --> 00:09:29.419 the agents check each other's work? It's AI peer 00:09:29.419 --> 00:09:32.480 review. It's formalized AI peer review. So if 00:09:32.480 --> 00:09:34.759 cursor writes the code, you immediately trigger 00:09:34.759 --> 00:09:37.360 Claude to review it, but with a different specialized 00:09:37.360 --> 00:09:39.799 prompt. You tell Claude its role is a strict 00:09:39.799 --> 00:09:43.210 senior tech lead. And the review prompt is the 00:09:43.210 --> 00:09:45.570 key here. You can't just say, look for errors. 00:09:45.690 --> 00:09:48.169 No, you have to give it critical evaluation criteria. 00:09:48.429 --> 00:09:50.490 Tell it to look for performance bottlenecks, 00:09:51.029 --> 00:09:54.230 like nested loops inside other loops, or check 00:09:54.230 --> 00:09:56.669 for obscure violations of your naming conventions. 00:09:56.970 --> 00:10:00.210 I saw a perfect example of this. An AI updated 00:10:00.210 --> 00:10:03.769 a config file correctly. But the reviewer, AI, 00:10:04.049 --> 00:10:06.649 the strict tech lead, commented that the main 00:10:06.649 --> 00:10:09.429 README file hadn't been updated to reflect the 00:10:09.429 --> 00:10:12.210 change. It found a communication gap a human 00:10:12.210 --> 00:10:14.909 might have easily missed. When you're using this 00:10:14.909 --> 00:10:17.509 triangle strategy, how do you stop the reviewer 00:10:17.509 --> 00:10:21.029 AI from just agreeing with the first AI to save 00:10:21.029 --> 00:10:24.559 time? This strict senior tech lead prompt enforces 00:10:24.559 --> 00:10:27.000 constructive but critical evaluation criteria, 00:10:27.139 --> 00:10:29.379 like performance metrics. It forces an adversarial 00:10:29.379 --> 00:10:32.039 role. This whole setup can sound complex, but 00:10:32.039 --> 00:10:34.039 the sources really emphasize you don't need to 00:10:34.039 --> 00:10:36.039 deploy all three bots at once. No, you start 00:10:36.039 --> 00:10:38.100 small. And we have a simple checklist for getting 00:10:38.100 --> 00:10:41.220 started. Okay, what's first? First, prepare your 00:10:41.220 --> 00:10:45.919 API keys. Claude, OpenAI, Cursor. A useful tip 00:10:45.919 --> 00:10:48.639 here is to use the Claude setup token command, 00:10:48.860 --> 00:10:50.759 which can tap into your monthly subscriptions 00:10:50.759 --> 00:10:53.879 and potentially save on costs. Second, and this 00:10:53.879 --> 00:10:56.960 is an absolute necessity, add those keys as secrets 00:10:56.960 --> 00:11:00.559 in your GitHub repository settings. Never, ever 00:11:00.559 --> 00:11:04.019 paste your keys directly into code files. Huge 00:11:04.019 --> 00:11:06.720 security failure waiting to happen. Right. And 00:11:06.720 --> 00:11:10.299 third. Start with the safest method. Create one 00:11:10.299 --> 00:11:13.019 simple hybrid workflow with Claude. Get that 00:11:13.019 --> 00:11:15.419 YAML file running requiring manual approval for 00:11:15.419 --> 00:11:17.659 everything until you've built up trust in the 00:11:17.659 --> 00:11:19.960 system. Finally, let's talk about common pitfalls. 00:11:20.179 --> 00:11:22.080 I learned some of these the hard way. The first 00:11:22.080 --> 00:11:25.399 is the infinite loop. Yes, a classic automation 00:11:25.399 --> 00:11:27.519 trap. You have to make sure the AI doesn't trigger 00:11:27.519 --> 00:11:30.000 itself. How does that happen? Well, if your workflow 00:11:30.000 --> 00:11:32.820 triggers on a new comment and the AI posts a 00:11:32.820 --> 00:11:34.639 comment like, I fixed it, that could trigger 00:11:34.639 --> 00:11:36.299 the workflow again. So you have to filter that 00:11:36.299 --> 00:11:38.480 out. Exactly. You check who the actor, the user 00:11:38.480 --> 00:11:40.600 of the comment is. If it's the bot, you break 00:11:40.600 --> 00:11:43.179 the loop. The second mistake is ignoring context. 00:11:44.039 --> 00:11:46.320 If you don't spend time writing a solidagents 00:11:46.320 --> 00:11:49.500 .md file, the AI is just going to guess at your 00:11:49.500 --> 00:11:52.179 project's style, and it will probably guess wrong. 00:11:52.559 --> 00:11:55.490 And finally, cost management. This is critical. 00:11:55.730 --> 00:11:58.210 Tokens are like water from a tap. Start with 00:11:58.210 --> 00:12:01.490 small, specific tasks like fix this one function 00:12:01.490 --> 00:12:04.549 in this one file. Don't say rewrite my entire 00:12:04.549 --> 00:12:06.909 app. You will get a very large bill if you do 00:12:06.909 --> 00:12:09.509 that. Be surgical in your requests. So what does 00:12:09.509 --> 00:12:11.889 this all mean for you, the listener? The core 00:12:11.889 --> 00:12:14.629 lesson here is this fundamental shift in your 00:12:14.629 --> 00:12:16.789 role. You go from user to manager. You aren't 00:12:16.789 --> 00:12:19.450 replacing yourself. You're directing a team of 00:12:19.450 --> 00:12:21.850 fast, tireless workers who handle the typing, 00:12:22.009 --> 00:12:23.750 the coding, the testing, even the first round 00:12:23.750 --> 00:12:26.389 of security checks, you get to focus on architecture 00:12:26.389 --> 00:12:28.830 and strategy. It really transforms the definition 00:12:28.830 --> 00:12:32.129 of a developer. Building this 2047 autonomous 00:12:32.129 --> 00:12:34.970 dev team feels like science fiction, but it's 00:12:34.970 --> 00:12:37.169 really just connecting existing tools through 00:12:37.169 --> 00:12:39.419 smart, repeatable workflows. And it leaves you 00:12:39.419 --> 00:12:41.159 with a really interesting question to think about. 00:12:41.559 --> 00:12:44.360 If these autonomous AI systems, driven by strict 00:12:44.360 --> 00:12:47.220 templates and validators like SonarKweeb, increasingly 00:12:47.220 --> 00:12:49.779 write and review our code based on fixed rules, 00:12:50.259 --> 00:12:52.519 will this ultimate focus on efficiency lead to 00:12:52.519 --> 00:12:55.299 a global homogenization of programming style? 00:12:55.620 --> 00:12:58.360 Will the unique creative quirks that human architects 00:12:58.360 --> 00:13:01.360 bring to software eventually fade away in favor 00:13:01.360 --> 00:13:04.139 of perfect, predictable, and maybe identical 00:13:04.139 --> 00:13:05.980 structure? Something to consider as you start 00:13:05.980 --> 00:13:07.179 building your own digital team.