0:00 I can't hear another word about agents. 0:03 This is the next big step forward. A 0:05 full coding agent. Every single software 0:07 company in the world needs an agentic 0:09 systems. 0:09 I mean agents is it's almost sort of a 0:11 buzz word, right? Like everybody you 0:13 talk to now is building agents. 0:14 Agentic coding experience. Do a 0:16 complicated coding agentic thing. 0:17 Codex is another sort of agent app. 0:19 What is an agent? 0:20 Agent agentic harness. 0:22 All right. Most explanations of agentic 0:24 AI systems are either too technical or 0:26 too basic. So this one's meant for 0:28 people like me. You're not a developer. 0:30 You use chat or claude every single day. 0:32 You've probably open Claude code or 0:34 co-work once or twice and weren't quite 0:36 sure what to do with it. And you want to 0:37 actually understand how this all fits 0:39 together without reading a single line 0:41 of code. Now, we're going to walk 0:42 through four levels. Chat bots, AI 0:45 workflows, aentic workflows, and aentic 0:48 AI systems. So, each one gives the AI 0:50 more autonomy, more capability, and at 0:53 each one, the wrapper around the AI gets 0:55 more sophisticated, too. And all of 0:57 those intimidating terms you see 0:58 everywhere, harness engineering, skills, 1:01 memory systems, MCPs, they are a lot 1:03 simpler than you think. And we'll knock 1:05 them down to plain English as we go 1:07 along. So I'm going to use one example 1:09 across all four levels so that you can 1:10 feel the difference at each level in the 1:13 work that you need to do. And that's 1:14 going to be content repurposing. So 1:15 we're going to take a YouTube video and 1:17 turn it into social posts, newsletters, 1:19 short form clips, the whole lot. So 1:22 let's see how that looks at each level 1:24 and where you can get more time back. So 1:26 level one is all about chat bots. Chat 1:28 GBT Claude Gemini. This is the one 1:30 you're probably familiar with, but 1:31 there's a few things that you need to 1:33 understand that are key to the higher 1:34 levels. So here's what content 1:36 repurposing looks like at level one. I 1:38 finish a recording of a YouTube video 1:40 like this one. I'll find a transcript. 1:42 I'm going to open Claude on the web and 1:43 I say, "Write me a LinkedIn post about 1:45 this." And Claude's going to come back. 1:47 It's going to write something. It's 1:48 fine. It reads like a LinkedIn post with 1:50 a ton of emojis, but it doesn't sound 1:52 like me. It doesn't know my audience. It 1:54 doesn't know that last month my 1:56 carousels outperformed my text posts and 1:58 it definitely doesn't know what I posted 2:00 last week. So, it might repeat the same 2:01 angle that I already used. Now, if I 2:03 wanted it to do better, I'd have to 2:04 paste in my voice guidelines, my 2:06 audience notes, my recent post history, 2:08 and I could do that all manually. But 2:10 obviously, I'm simplifying a bit. Claude 2:12 and GPT have projects. Gemini has gems, 2:15 and you can actually store static 2:16 contacts there. But ultimately, that's 2:18 what it is. It's static context. So, 2:20 there's two things you need to remember 2:21 about chat bots. One, they don't know 2:23 your business, none of your brand, your 2:25 audience, the content history of your 2:28 last few posts. And two, they're 2:30 passive. They wait for you to prompt 2:31 them. They don't go and do anything on 2:33 their own. So, at level one, the chatbot 2:35 is there giving you advice, but it's not 2:37 executing. So, keep that in mind. Now, 2:39 level two was all the rave in 2025. AI 2:42 workflow. So, think n, zapia, and 2:44 make.com. Everyone was talking about 2:46 this, including me in 2025. And here's 2:49 what content repurposing looks like at 2:50 level two. So, I build out an N workflow 2:53 and every time I publish a new YouTube 2:54 video, the workflow is going to fire 2:56 automatically. So, step one is going to 2:58 pull the transcript. Step two, it's 3:00 going to send that to Claude through the 3:02 AI node with my voice guidelines 3:04 hardcoded into the prompt. Step three, 3:06 Claude is going to come back with a 3:07 draft LinkedIn post. And step four, the 3:10 workflow is going to drop it into my 3:11 scheduling tool so I can just review the 3:13 draft. And for somebody that's come from 3:14 just working with ChatVt or Claude, this 3:16 is going to feel like magic. So you can 3:18 get it to work on repeat using the same 3:20 steps every single time. So I've gone 3:22 from manually pasting my transcript into 3:24 a chat window to having a draft appear 3:26 in myuler without lifting a finger. Here 3:29 is the limitation though. The workflow 3:31 can't think. If my best performing posts 3:33 this month have been carousels and not 3:35 text posts, the workflow doesn't know 3:37 that. It's got my static prompt 3:38 templates that I put in there 3 or 4 3:41 months ago. I can of course go and 3:42 update that manually, but it's not 3:44 automatically adapting. How about if the 3:45 video topic doesn't even suit LinkedIn 3:47 at all and would be better as a Twitter 3:49 thread? The workflow can't make that 3:51 judgment call because it's just running 3:53 through the same steps in the same order 3:55 regardless. It's not like having an 3:57 employee who can do the thinking. And 3:58 ultimately, if the output's not good 4:00 enough, I'm going to go back in and 4:01 rewrite the prompts myself. So, at level 4:03 two, the AI is doing some of the work, 4:05 but it's not making decisions. It's just 4:07 following our exact step-by-step guide. 4:09 And those are the steps that you've 4:10 defined for it. Now, the next part is 4:11 the most important concept in this 4:13 entire video. The difference between an 4:15 AI workflow and an agentic workflow at 4:18 level three is who decides the execution 4:21 path. So at level two, you've defined 4:23 the steps. At level three, the model 4:25 does, and you're free to do some other 4:27 task while it's actually making those 4:29 decisions. But before we move on to 4:31 that, YouTube tells me that 97% of you 4:33 watching this video right now haven't 4:35 subscribe to the channel. So do me a 4:37 quick favor and hit the subscribe button 4:38 below if you've made it this far. So 4:40 level three then is agentic workflows. 4:43 And here's what content repurposing 4:44 looks like at level three. I open Claude 4:46 Code and say, "Turn this week's video 4:48 into content for LinkedIn, Twitter, and 4:50 my newsletter." And Claude Code is going 4:53 to actually just go and figure it all 4:54 out. So, it pulls the transcript. It 4:56 reads my brand voice file. It's going to 4:58 look at the video topic and decide which 5:00 moments in that transcript of the video 5:02 work best for each platform based on 5:04 different viral trend indicators. It's 5:07 going to draft a LinkedIn carousel 5:08 because the topic suits visual 5:10 storytelling. It might write an X thread 5:12 because there's a strong contrarian 5:14 angle and X loves contrarian. It's then 5:16 going to run all of those through my 5:17 style guide through my brand voice and 5:19 rewrite the ones that don't pass the 5:21 criteria. Everything is then saved for 5:23 me to review. So, I didn't write those 5:25 steps. The model decided them based on 5:28 the goal that I gave it. And quite 5:29 frankly, this is what it would look like 5:31 if we did this kind of workflow inside 5:33 N. It would have that infinite loop 5:35 logic and it would be a mess to 5:36 maintain. Trust me, I'm talking from 5:39 experience on building these. And the 5:40 technical name for this agent loop is 5:42 actually react. So it's reason and act. 5:46 The model reasons about what to do, acts 5:48 on it, observe the result, and iterates 5:51 until it's done. So finally, we can 5:53 start getting our time back because the 5:55 model's not only doing the work, it's 5:57 also now doing the thinking. Now the 5:59 tools at this level are things like claw 6:00 code, open's codeex, and cursor. And 6:04 what they all have in common is they're 6:05 what's called a harness. So a harness is 6:08 the infrastructure that surrounds the 6:09 model to make it more reliable, 6:12 controllable, and deployable for actual 6:14 business use cases. So it's what turns 6:16 thinking into doing. So without a 6:18 harness, you've got a chatbot in a 6:19 browser tab. With a harness, the model 6:21 can read your files, run commands, call 6:24 other tools, and check its own work. 6:26 There are lots of other things that you 6:28 can add onto a harness, but you can't 6:30 add them onto a chatbot. And we'll come 6:31 back to that later. So claw code is 6:33 actually just a harness. Codeex is a 6:35 harness. cursor is a harness. It's 6:37 different products, but the exact same 6:39 idea. They wrap around the model and 6:40 give it the ability to act on your 6:42 computer with your files using your 6:44 tools. So, at level three, you've got a 6:46 single agent running in a harness 6:48 working toward one goal. And it's 6:50 powerful. And this is good enough for 6:52 most people, but it does top out. One 6:54 agent with one goal in one terminal 6:56 session. It doesn't remember what it 6:58 learned last week. It doesn't know which 7:00 of my posts are performing best at the 7:02 moment. And if I wanted to also extract 7:03 video clips, build carousels, generate 7:05 some ad copy maybe, and schedule 7:08 everything, I'd have to run each task 7:09 separately and reexplain some of the 7:11 context every time. So if level three is 7:13 one agent on one task, then level four 7:15 is a coordinated team running your 7:18 operations. So this is all about agentic 7:21 AI systems. And here is what content 7:23 repurposing looks like at level four. So 7:25 I trigger one command by just describing 7:27 what I want and the full content engine 7:29 runs from one YouTube video. So, there's 7:31 going to be a skill in the background 7:32 that extracts the best clips from the 7:34 video for short form content and ranks 7:36 them on different criteria. There's 7:38 going to be another one that builds 7:39 platform specific carousels with the 7:41 right dimensions, the visual aesthetics 7:43 that are going to match my brand and 7:45 copy for each different platform. 7:47 Another instance is going to draft the 7:48 weekly newsletter from key takeaways. 7:51 Another is going to generate ad copy 7:53 from the angles that performed best in 7:55 the past. And finally, everything is 7:56 going to come together and be cued into 7:58 my scheduling tool. So each skill on its 8:00 own has its own instructions, its own 8:02 quality bar, its own output format, and 8:04 the system as it goes through is going 8:06 to load my brand context in at the right 8:08 time and only the sections or context 8:10 references that are actually needed at 8:12 that point. It's then able to actually 8:13 go and check its own work, flag anything 8:15 that needs me and handle the rest 8:17 autonomously by itself. So you can think 8:18 of this instead of one agent doing one 8:20 job, you've got a system running entire 8:22 content operation on your behalf. So the 8:25 jump then from level three to level four 8:26 is what people are actually calling 8:28 harness engineering. And the building 8:30 blocks are simpler than they sound. So 8:32 you've got skills which are just folders 8:34 of instructions for specific tasks like 8:36 how to write a LinkedIn carousel, how to 8:39 extract short form clips from a long 8:40 form transcript, how to draft a 8:42 newsletter in my voice, and the agent is 8:44 going to load the right skill only when 8:46 it needs it, and the right reference 8:48 examples only when it needs it, too. So 8:50 it's not going to bloat your context or 8:52 cost you a bunch of tokens. Then we 8:54 connect up to our tools with MCPs or 8:56 model context protocol. So the 8:57 scheduling platform, the analytics 8:59 dashboard, maybe your CRM. MCP is just 9:02 the standard for plugging things in to 9:04 our claude code harness. We've got one 9:06 of the most important concepts, memory. 9:07 So the system carries context between 9:09 sessions, so it knows which post 9:11 performed best last month. It remembers 9:13 which newsletter subject lines got the 9:15 highest open rates. It builds on what 9:17 it's learned instead of starting from 9:18 zero every time. And memory can 9:20 literally be as simple as a markdown 9:21 file that the system is going to read 9:23 and update. Or you can go more complex 9:25 with a sophisticated database that 9:27 connects across all your different LLMs 9:29 and tools. So it almost sounds too good 9:30 to be true and you're probably thinking, 9:32 "So this thing is just going to run on 9:34 its own and do all my work. I'm going to 9:35 need no employees." Well, no, not 9:38 exactly. By design, we've got human in 9:40 the loop. So set it and forget it is not 9:43 a design principle that we've got here. 9:44 the systems that actually work and I've 9:46 seen work in the real world have a human 9:48 review built in at the right places. So 9:50 when you're designing these systems, you 9:52 have to be deliberate about where you 9:53 sit in the loop or where somebody who's 9:55 engaging with the system is going to get 9:57 feedback or give feedback in the loop. 9:58 And usually that's in either the input 10:00 stage or the output stage where we're 10:02 doing a quality review so that we can 10:04 make sure it's good before it goes live. 10:05 So in my content system, for example, 10:07 the AI does all the heavy lifting. is 10:09 going to draft, check, and format as per 10:11 my rules, but nothing's going to get 10:13 published without me seeing it first. I 10:15 want to make sure that it meets my human 10:16 criteria before publishing. I've not yet 10:19 come across a system that's powerful 10:21 enough and knows me well enough to get 10:24 past this human in the loop stage. We 10:26 can get 95% of the way there with this 10:28 Agentic AI system. So, the tools at this 10:31 level are not things like claw code out 10:33 of the box anymore. are things like the 10:35 agentic operating system which is what 10:37 I'm building inside my own community and 10:39 open claw or Hermes so it's the 10:41 open-source personal agent that is 10:43 absolutely skyrocketed this year both 10:45 are taking the exact same approach build 10:47 a richer system of files on top of the 10:50 base agent so it can handle actual real 10:52 complexity and feed in context at the 10:54 right time with a human in the seat with 10:56 a bit of control and here's the bit that 10:58 nobody actually mentions the bit where 11:00 you can probably at this point breathe a 11:02 sigh of relief it's not as complicated 11:04 ated as it sounds because underneath all 11:06 this is just files in folders. So the 11:08 fancy terminology is pretty much just 11:11 the front and what's really happening is 11:13 we're giving the model access to the 11:15 right file in the right folder at the 11:18 right time. That is simply it. So a 11:20 skill for example is just a markdown 11:22 file. The brand voice, the quality 11:24 rules, the memory, all readable and 11:27 editable directly in your files. So 11:29 organizing something like this is 11:30 actually closer to building a notion 11:32 workspace out than it is to writing 11:35 code. Which means that this isn't just 11:36 for developers. And the audience for 11:38 aentic AI systems is actually business 11:40 owners, knowledge workers, people who 11:42 want their operation to actually run on 11:44 AI without learning to code. So to wrap 11:46 up, we've got four levels. Level one is 11:48 chat bots. You're going to pay something 11:50 in, get text back, but it's all advice 11:52 and no action. Then we hit AI workflow. 11:54 So you're building out a pipeline in an 11:56 automation tool and the AI is going to 11:58 fill in some of the gaps, but it's going 11:59 to follow the same steps in the same 12:01 order every time. We want it to be a 12:03 little bit more autonomous in level 12:04 three where we hand it off to an agentic 12:06 workflow. So you're basically giving it 12:08 a goal and the agent is going to 12:09 determine the steps with your context. 12:11 And that's when we wrap it in a harness 12:13 like cloud code. But if you really want 12:15 to achieve production grade results, 12:17 things that are going to run 95% without 12:19 you, then you need to develop agentic AI 12:21 systems where you've got multiple 12:22 skills, a shared memory coordinated by 12:25 the system with you in the loop exactly 12:28 where it counts. And we've distilled 12:29 this to just folders and files on top of 12:32 something like claw code and codeex. So 12:34 if you found this helpful, you might 12:36 want to learn how to build your own 12:37 level 4 aentic AI system. So, we'll show 12:40 you how we made ours in the next