WEBVTT 00:00:00.000 --> 00:00:03.120 We treat AI like a coding vending machine. You 00:00:03.120 --> 00:00:05.700 punch a prompt in, you get a snippet out. But 00:00:05.700 --> 00:00:08.220 that approach is actually completely wrong. Yeah, 00:00:08.259 --> 00:00:10.660 it really is. The future is building a custom 00:00:10.660 --> 00:00:13.359 coding environment. It needs to deeply understand 00:00:13.359 --> 00:00:17.879 how you actually work. Welcome to the Deep Dive. 00:00:18.719 --> 00:00:21.260 Today we're exploring a massive shift in software 00:00:21.260 --> 00:00:23.899 engineering. We're looking closely at DeepSeq 00:00:23.899 --> 00:00:26.820 Harness today. It's emerged as a very powerful 00:00:26.820 --> 00:00:30.719 new technology. It's essentially a free open 00:00:30.719 --> 00:00:34.299 source alternative to Claude code. Right. We're 00:00:34.299 --> 00:00:36.060 going to unpack what an agent harness actually 00:00:36.060 --> 00:00:38.780 is. Then we'll explore the architectural hurdles 00:00:38.780 --> 00:00:41.399 of local setup. We aren't just reading terminal 00:00:41.399 --> 00:00:43.560 commands today. No, definitely not. We're looking 00:00:43.560 --> 00:00:46.140 at why local sovereign environments matter. We'll 00:00:46.140 --> 00:00:48.420 also explore building custom memory tools together. 00:00:48.679 --> 00:00:50.799 This is where the workflow completely changes. 00:00:51.060 --> 00:00:52.820 Finally, we'll weigh it against Cloud Code directly. 00:00:52.939 --> 00:00:54.799 We really need to see which approach actually 00:00:54.799 --> 00:00:56.659 wins out. Yeah, it's a great comparison. Let's 00:00:56.659 --> 00:00:59.259 start by defining the core technology here. There 00:00:59.259 --> 00:01:01.460 is a fundamental paradigm shift happening right 00:01:01.460 --> 00:01:04.420 now. There really is. Developers are moving away 00:01:04.420 --> 00:01:07.299 from simple AI chat models. They're moving toward 00:01:07.299 --> 00:01:10.819 complex, personalized agent harnesses instead. 00:01:11.219 --> 00:01:13.099 We really need to define those two terms clearly. 00:01:13.840 --> 00:01:16.420 What exactly is an AI model in this context? 00:01:16.620 --> 00:01:20.040 An AI model handles the raw thinking and processing. 00:01:20.299 --> 00:01:22.840 It understands your requests and generates the 00:01:22.840 --> 00:01:25.099 actual text responses. It's the brain inside 00:01:25.099 --> 00:01:28.019 the software system. Yeah, exactly. But an agent 00:01:28.019 --> 00:01:30.420 harness is a very different thing. How would 00:01:30.420 --> 00:01:32.980 you define an agent harness simply? A framework 00:01:32.980 --> 00:01:36.340 that helps an AI interact with your actual tools. 00:01:36.599 --> 00:01:37.760 That makes a lot of sense when you think about 00:01:37.760 --> 00:01:40.680 it. The harness is the hands and eyes of the 00:01:40.680 --> 00:01:43.700 system. DeepSeq Harness focuses entirely on that 00:01:43.700 --> 00:01:46.060 second structural part. Right. It provides a 00:01:46.060 --> 00:01:49.019 flexible, plug -in -based architecture for the 00:01:49.019 --> 00:01:51.920 AI model. Developers can extend it for many different 00:01:51.920 --> 00:01:55.140 custom coding workflows. So why are developers 00:01:55.140 --> 00:01:57.640 flocking to this specific tool right now? Well, 00:01:57.900 --> 00:01:59.959 the first major reason is that it's completely 00:01:59.959 --> 00:02:02.980 open source. Developers love being able to explore 00:02:02.980 --> 00:02:05.670 the raw codebase themselves. Oh, absolutely. 00:02:05.829 --> 00:02:08.750 They can customize DeepSeq Harness based on their 00:02:08.750 --> 00:02:11.889 exact daily requirements. Another huge factor 00:02:11.889 --> 00:02:14.610 is that it runs entirely locally. Yeah, that's 00:02:14.610 --> 00:02:17.030 huge. This touches on a massive debate in software 00:02:17.030 --> 00:02:19.689 engineering right now. Developers want to tightly 00:02:19.689 --> 00:02:22.150 control their own secure coding environment. 00:02:22.490 --> 00:02:25.050 They manage exactly how the system runs on their 00:02:25.050 --> 00:02:28.490 own machines. No proprietary code is being sent 00:02:28.490 --> 00:02:31.349 to a third -party cloud server. It also supports 00:02:31.349 --> 00:02:33.550 multiple different AI models right out of the 00:02:33.550 --> 00:02:36.110 box. Right. You can choose different models depending 00:02:36.110 --> 00:02:38.909 on your specific daily tasks. And it allows for 00:02:38.909 --> 00:02:41.250 custom tools and memory workflows to be built. 00:02:41.469 --> 00:02:44.169 You can create workflows that match your actual 00:02:44.169 --> 00:02:47.009 human development process. Flexibility is definitely 00:02:47.009 --> 00:02:49.289 the biggest advantage of this entire ecosystem. 00:02:49.469 --> 00:02:51.889 Using a fixed AI assistant is like buying a pre 00:02:51.889 --> 00:02:54.240 -built gaming console. Oh, I like that. You plug 00:02:54.240 --> 00:02:56.379 it in and you're stuck with its lock software. 00:02:57.219 --> 00:02:59.560 But DeepSeq Harness is like building a custom 00:02:59.560 --> 00:03:03.240 PC from scratch. You choose every single component 00:03:03.240 --> 00:03:05.740 to fit your exact performance needs. That's a 00:03:05.740 --> 00:03:07.620 really great way to visualize the core difference. 00:03:08.300 --> 00:03:11.020 You start with a simple AI coding assistant first. 00:03:11.680 --> 00:03:13.939 Then you gradually build a more personalized, 00:03:14.419 --> 00:03:16.939 complex development environment. But is building 00:03:16.939 --> 00:03:19.520 a harness just reinventing the wheel for the 00:03:19.520 --> 00:03:22.870 sake of tinkering? Not at all. A fixed tool forces 00:03:22.870 --> 00:03:26.289 you to adapt to its invisible limitations. A 00:03:26.289 --> 00:03:29.409 harness gives you absolute long -term workflow 00:03:29.409 --> 00:03:32.389 flexibility. You build it once, and it scales 00:03:32.389 --> 00:03:35.289 with your specific evolving needs. So it's a 00:03:35.289 --> 00:03:37.669 customizable workspace, not just a fixed tool. 00:03:37.909 --> 00:03:40.310 Exactly. But because it's a custom environment, 00:03:40.669 --> 00:03:42.909 you must actively build it. Let's look at the 00:03:42.909 --> 00:03:45.430 actual installation and setup philosophy. You 00:03:45.430 --> 00:03:47.469 have to prepare your local computer environment 00:03:47.469 --> 00:03:49.479 first. We aren't going to read line -by -line 00:03:49.479 --> 00:03:51.659 terminal commands here. The audience knows how 00:03:51.659 --> 00:03:53.979 to use a command line interface. Right. But we 00:03:53.979 --> 00:03:56.280 should discuss the architectural hurdles involved 00:03:56.280 --> 00:03:58.860 in this setup. You have to build a solid local 00:03:58.860 --> 00:04:01.719 scaffolding first. You need foundational tools 00:04:01.719 --> 00:04:04.199 installed securely on your local machine. You're 00:04:04.199 --> 00:04:06.400 relying on core package managers to handle the 00:04:06.400 --> 00:04:08.560 heavy lifting. You'll check for Node .js and 00:04:08.560 --> 00:04:12.020 Git first. Node .js is a runtime for executing 00:04:12.020 --> 00:04:14.439 JavaScript code locally on your machine. Yeah. 00:04:14.699 --> 00:04:17.180 And you also need to install the PNPM package 00:04:17.180 --> 00:04:20.180 manager. You are essentially establishing a sovereign 00:04:20.180 --> 00:04:22.939 development environment on your laptop. It isn't 00:04:22.939 --> 00:04:25.959 a simple, consumer -friendly, one -click install 00:04:25.959 --> 00:04:28.199 process right now. No, it's definitely not. You 00:04:28.199 --> 00:04:30.819 are literally pulling down the raw repository 00:04:30.819 --> 00:04:33.300 source code. You're compiling it directly using 00:04:33.300 --> 00:04:35.720 local web development package managers. After 00:04:35.720 --> 00:04:38.680 pulling down the repository, you build the actual 00:04:38.680 --> 00:04:42.459 harness. Beat. Then you launch a local web UI 00:04:42.459 --> 00:04:45.040 on your own machine. This interface is where 00:04:45.040 --> 00:04:47.279 you work with your new coding agent. But after 00:04:47.279 --> 00:04:49.379 getting the environment running, you must choose 00:04:49.379 --> 00:04:52.439 an AI model. Right. You open the model selector 00:04:52.439 --> 00:04:55.480 inside your newly hosted local web UA. You choose 00:04:55.480 --> 00:04:57.939 a model based strictly on your specific coding 00:04:57.939 --> 00:05:01.579 needs. DeepSeek V4 Pro is designed specifically 00:05:01.579 --> 00:05:05.040 for highly complex coding tasks. It provides 00:05:05.040 --> 00:05:07.259 much deeper reasoning for extremely difficult 00:05:07.259 --> 00:05:09.519 software engineering problems. On the other hand, 00:05:09.740 --> 00:05:12.399 DeepSeek V4 Flash provides significantly faster 00:05:12.399 --> 00:05:15.300 responses overall. It's much better for simpler, 00:05:15.740 --> 00:05:18.410 everyday, repetitive coding tasks. The selected 00:05:18.410 --> 00:05:21.649 model handles understanding requests and analyzing 00:05:21.649 --> 00:05:24.810 complex logic problems. Meanwhile, the harness 00:05:24.810 --> 00:05:26.730 provides the structured environment where the 00:05:26.730 --> 00:05:29.269 agent actually works. Next, you have to set up 00:05:29.269 --> 00:05:32.060 your core agent prompt. This is where the system 00:05:32.060 --> 00:05:34.639 gets really interesting to me. You give it strict 00:05:34.639 --> 00:05:37.160 instructions for handling your daily coding tasks. 00:05:37.600 --> 00:05:40.420 You tell the AI to analyze the entire project 00:05:40.420 --> 00:05:43.420 structure first. It must identify root causes 00:05:43.420 --> 00:05:46.160 before suggesting any actual code changes. It 00:05:46.160 --> 00:05:47.819 also needs to explain its technical approach 00:05:47.819 --> 00:05:50.399 clearly beforehand. It should only update the 00:05:50.399 --> 00:05:53.199 absolutely necessary project files. It must strictly 00:05:53.199 --> 00:05:56.240 avoid making any unnecessary formatting changes 00:05:56.240 --> 00:05:59.149 to your code. Finally, it must review the final 00:05:59.149 --> 00:06:02.509 changes for potential hidden issues. It has to 00:06:02.509 --> 00:06:05.189 keep the code base clean and highly maintainable 00:06:05.189 --> 00:06:07.610 over time. It must stay perfectly consistent 00:06:07.610 --> 00:06:10.170 with the existing project coding style. When 00:06:10.170 --> 00:06:12.829 completing a task, it must summarize the specific 00:06:12.829 --> 00:06:16.069 changes made. It has to provide the exact reasoning 00:06:16.069 --> 00:06:18.870 behind each specific file edit. It should also 00:06:18.870 --> 00:06:20.910 list possible architectural improvements you 00:06:20.910 --> 00:06:23.730 should carefully consider. Eat, eat. I still 00:06:23.730 --> 00:06:26.089 wrestle with prompt drift myself. Oh, it's very 00:06:26.089 --> 00:06:28.910 common. The AI just slowly forgets your original 00:06:28.910 --> 00:06:31.350 rules over time. You write perfect instructions 00:06:31.350 --> 00:06:33.550 for the agent to follow closely. Three hours 00:06:33.550 --> 00:06:36.629 later, it's suddenly hallucinating entirely new 00:06:36.629 --> 00:06:38.949 formatting rules. Right, and that's why these 00:06:38.949 --> 00:06:42.649 rigid system guardrails are so vital. This highly 00:06:42.649 --> 00:06:45.230 structured prompt helps create much more consistent 00:06:45.230 --> 00:06:48.029 long -term results. What happens if you skip 00:06:48.029 --> 00:06:50.930 these strict prompt guardrails entirely? The 00:06:50.930 --> 00:06:55.050 AI might make unnecessary, messy changes across 00:06:55.050 --> 00:06:57.790 various disconnected files. It could break working 00:06:57.790 --> 00:06:59.810 code without you even realizing it happened. 00:06:59.930 --> 00:07:01.490 You give it strict rules so it doesn't break 00:07:01.490 --> 00:07:04.189 things. Exactly. And that brings us to the next 00:07:04.189 --> 00:07:06.610 massive conceptual phase. Getting it running 00:07:06.610 --> 00:07:08.889 locally is really just step one of the journey. 00:07:09.029 --> 00:07:11.430 The real power unlocks when you give it long 00:07:11.430 --> 00:07:13.730 -term structural memory. You can customize how 00:07:13.730 --> 00:07:16.009 your agent understands whole complex software 00:07:16.009 --> 00:07:18.269 projects. DeepSeq Harness allows you to build 00:07:18.269 --> 00:07:21.019 completely custom, specialized workflow tools. 00:07:21.240 --> 00:07:24.360 You can create reusable utility tools for specific 00:07:24.360 --> 00:07:27.139 software frameworks. It works great with popular 00:07:27.139 --> 00:07:30.160 front -end web frameworks like React. It also 00:07:30.160 --> 00:07:33.220 supports modern back -end API frameworks like 00:07:33.220 --> 00:07:36.839 Fast API beautifully. You can build robust tools 00:07:36.839 --> 00:07:39.800 for scripting languages like JavaScript or Python. 00:07:39.959 --> 00:07:42.759 It even supports complex systems programming 00:07:42.759 --> 00:07:46.040 languages like Go Perfectly. These custom tools 00:07:46.040 --> 00:07:49.560 help the AI analyze your specific code base structure. 00:07:49.709 --> 00:07:52.410 They can thoroughly review files and check project 00:07:52.410 --> 00:07:55.410 structures constantly. They handle tedious, repetitive 00:07:55.410 --> 00:07:58.209 development tasks without needing constant human 00:07:58.209 --> 00:08:00.850 input. But you must define a clear purpose for 00:08:00.850 --> 00:08:03.230 every single custom tool. Right. You have to 00:08:03.230 --> 00:08:05.649 establish the required data inputs and the expected 00:08:05.649 --> 00:08:07.970 outputs. You also need to outline the specific 00:08:07.970 --> 00:08:10.310 workflow steps involved internally. Then you 00:08:10.310 --> 00:08:12.829 define the exact implementation approach for 00:08:12.829 --> 00:08:14.970 the new tool. Building this memory system is 00:08:14.970 --> 00:08:17.670 like stacking Lego blocks of data. Oh yeah! You 00:08:17.670 --> 00:08:20.509 assemble individual, simple pieces into a complex, 00:08:20.790 --> 00:08:22.629 highly intelligent structure. You can create 00:08:22.629 --> 00:08:25.089 advanced memory systems and deeply personal workflows 00:08:25.089 --> 00:08:27.610 easily. These systems store important project 00:08:27.610 --> 00:08:30.410 information and specific coding preferences safely. 00:08:30.670 --> 00:08:33.429 They remember complex architectural rules that 00:08:33.429 --> 00:08:36.070 the AI can use later. This directly addresses 00:08:36.070 --> 00:08:38.789 the massive cognitive load of modern AI coding. 00:08:39.200 --> 00:08:42.340 Think about how exhausting current chat -based 00:08:42.340 --> 00:08:45.299 AI interfaces actually are. It's a lot. You have 00:08:45.299 --> 00:08:47.720 to explain your entire project architecture every 00:08:47.720 --> 00:08:49.679 single time. You open a fresh window and you're 00:08:49.679 --> 00:08:52.240 starting from absolute zero. But with this harness, 00:08:52.600 --> 00:08:55.559 the AI simply remembers your previous context. 00:08:55.919 --> 00:08:58.980 There are five key conceptual steps to this advanced 00:08:58.980 --> 00:09:01.320 memory system. We need to look at how they actually 00:09:01.320 --> 00:09:04.259 reduce developer cognitive load. Step one is 00:09:04.259 --> 00:09:07.460 focused entirely on comprehensive project understanding. 00:09:07.679 --> 00:09:10.480 Think about inheriting a massive, undocumented, 00:09:10.539 --> 00:09:13.899 messy legacy code base yourself. The AI learns 00:09:13.899 --> 00:09:16.340 how to review that completely new software project 00:09:16.490 --> 00:09:18.690 automatically. It knows exactly what critical 00:09:18.690 --> 00:09:20.830 architectural information it should review first. 00:09:21.110 --> 00:09:23.470 Step two is building the actual memory system 00:09:23.470 --> 00:09:26.049 structure itself. It dictates how project context 00:09:26.049 --> 00:09:28.409 and developer preferences are properly stored 00:09:28.409 --> 00:09:31.129 locally. It reliably stores previous coding decisions 00:09:31.129 --> 00:09:34.529 for future automated background tasks. Step three 00:09:34.529 --> 00:09:37.570 involves a highly detailed task analysis process. 00:09:38.009 --> 00:09:41.990 The AI analyzes new coding requests very methodically 00:09:41.990 --> 00:09:45.039 every single time. It identifies affected files 00:09:45.039 --> 00:09:47.440 and plans thoroughly before making any actual 00:09:47.440 --> 00:09:50.480 changes. Step 4 tackles the actual code changes 00:09:50.480 --> 00:09:53.480 and review process. The AI explains its proposed 00:09:53.480 --> 00:09:56.019 structural changes before editing any source 00:09:56.019 --> 00:09:58.519 files. Then it reviews the final technical result 00:09:58.519 --> 00:10:01.220 for potential hidden software issues. Step 5 00:10:01.220 --> 00:10:03.779 is strict coding standards management for your 00:10:03.779 --> 00:10:07.080 entire engineering team. The AI maintains strict, 00:10:07.460 --> 00:10:10.559 unwavering consistency with your specific formatting 00:10:10.559 --> 00:10:13.259 standards. It meticulously follows your project 00:10:13.259 --> 00:10:15.720 -specific rules without constantly being reminded. 00:10:15.919 --> 00:10:18.720 Two -sec silence. It completely changes how a 00:10:18.720 --> 00:10:21.220 developer approaches a massive legacy codebase. 00:10:21.500 --> 00:10:25.250 Whoa! Imagine an AI perfectly memorizing your 00:10:25.250 --> 00:10:27.769 entire enterprise architecture. It really is 00:10:27.769 --> 00:10:29.750 a staggering shift in how we write software. 00:10:30.190 --> 00:10:32.330 The workflow adapts entirely to your specific 00:10:32.330 --> 00:10:34.570 human software development process. It reliably 00:10:34.570 --> 00:10:37.129 follows your complex project rules and preferred 00:10:37.129 --> 00:10:39.509 coding style. How does this memory actually speed 00:10:39.509 --> 00:10:41.730 up daily work compared to a fresh chat window? 00:10:41.929 --> 00:10:44.669 Well, a fresh chat requires endless context setting 00:10:44.669 --> 00:10:47.769 every single time you start. Memory provides 00:10:47.769 --> 00:10:50.870 compounding time savings because the AI already 00:10:50.870 --> 00:10:53.350 knows your logic. It basically learns your habits 00:10:53.350 --> 00:10:55.970 and adapts to your style. Exactly. It stops being 00:10:55.970 --> 00:10:58.029 a standalone tool and becomes a true environment. 00:10:58.309 --> 00:11:00.950 If DeepSeq Harness is this powerful, why would 00:11:00.950 --> 00:11:03.850 anyone still use Claude code? They are simply 00:11:03.850 --> 00:11:06.389 designed for very different types of software 00:11:06.389 --> 00:11:09.529 engineering users. Let's compare them in a direct 00:11:09.529 --> 00:11:11.929 head -to -head technical showdown. Claude Code 00:11:11.929 --> 00:11:14.850 provides a highly stable, instantly ready -to 00:11:14.850 --> 00:11:17.669 -use coding experience. It gives incredibly consistent 00:11:17.669 --> 00:11:20.049 results for strict production -level software 00:11:20.049 --> 00:11:22.570 code. It is best for developers who want absolutely 00:11:22.570 --> 00:11:25.149 zero initial setup time. They just want to focus 00:11:25.149 --> 00:11:27.710 entirely on building consumer software products 00:11:27.710 --> 00:11:30.169 quickly. Cloud provides a ready -to -use cloud 00:11:30.169 --> 00:11:32.649 workflow with much less configuration required. 00:11:32.950 --> 00:11:35.210 It is a highly reliable assistant for important, 00:11:35.470 --> 00:11:37.970 fast -paced client -facing projects. But DeepSeek 00:11:37.970 --> 00:11:39.929 Harness is much better for developers wanting 00:11:39.929 --> 00:11:42.690 deep local control. It is designed for enterprise 00:11:42.690 --> 00:11:45.129 users who want to aggressively optimize software 00:11:45.129 --> 00:11:48.909 costs. It absolutely requires significantly more 00:11:48.909 --> 00:11:51.690 initial configuration to build the custom environment. 00:11:51.929 --> 00:11:55.309 But it allows you to create fully custom deeply 00:11:55.309 --> 00:11:57.950 integrated software tools. You build an AI environment 00:11:57.950 --> 00:12:00.330 that matches your exact personal engineering 00:12:00.330 --> 00:12:04.009 process. Is there a clear definitive winner between 00:12:04.009 --> 00:12:06.429 these two different approaches? It depends purely 00:12:06.429 --> 00:12:09.070 on what you value more in your workflow right 00:12:09.070 --> 00:12:11.509 now. You either want out -of -the -box convenience 00:12:11.509 --> 00:12:14.330 or you want total sovereign system control. Claude 00:12:14.330 --> 00:12:16.929 is for convenience. DeepSeek is for total control. 00:12:17.029 --> 00:12:19.490 Yeah. Sponsor. Let's step back and look at the 00:12:19.490 --> 00:12:22.799 big picture of this industry shift. The era of 00:12:22.799 --> 00:12:25.940 simple chat -based AI assistance is ending very 00:12:25.940 --> 00:12:28.360 quickly. We're moving far past just asking a 00:12:28.360 --> 00:12:30.879 cloud chatbot simple programming questions. The 00:12:30.879 --> 00:12:33.500 future definitely belongs to highly personalized, 00:12:33.879 --> 00:12:36.860 local AI coding environments. Developers will 00:12:36.860 --> 00:12:39.279 increasingly rely on bespoke automated workflows 00:12:39.279 --> 00:12:41.580 for their daily projects. The digital environment 00:12:41.580 --> 00:12:44.100 will finally adapt to the individual human developer. 00:12:44.200 --> 00:12:46.159 You no longer have to adapt your workflow to 00:12:46.159 --> 00:12:49.259 a fixed AI chatbot. You combine the right underlying 00:12:49.259 --> 00:12:52.559 AI model with a well -designed personalized workflow. 00:12:52.720 --> 00:12:54.940 You create a virtual workspace that fits exactly 00:12:54.940 --> 00:12:57.960 how you build complex software. This massive 00:12:57.960 --> 00:13:00.740 shift toward total personalization raises a truly 00:13:00.740 --> 00:13:03.919 fascinating philosophical question. If every 00:13:03.919 --> 00:13:06.759 developer builds their own highly idiosyncratic 00:13:06.759 --> 00:13:10.320 AI coding environment, Geet, an environment that 00:13:10.320 --> 00:13:12.519 completely memorizes their specific formatting 00:13:12.519 --> 00:13:15.840 quirks and mental shortcuts, Beat, will human 00:13:15.840 --> 00:13:18.240 engineers eventually lose the ability to seamlessly 00:13:18.240 --> 00:13:20.909 jump into a colleague's workflow? Thank you for 00:13:20.909 --> 00:13:22.509 listening. Keep exploring the future of code.