Dynamic Workflows are crazy. This is how to actually use them (beginner's guide)

Opus 4.8 just dropped with a bunch of new features including dynamic workflows, which are absolutely crazy. So in this video, we're going to run through them together and I can show you exactly how they work and what they can actually do. So just starting off with the release notes from Anthropic, basically the new features— of course, 4.8 has just launched, and then alongside that, it's also allowed us to adjust the effort levels inside of Claude Chat, which we've had inside of Claude Code for a little while, but it's nice to now have that inside of the Claude Chat. And then also in Claude Code, we have the new dynamic workflows, which of course we're going to be running through in this video. And then as well as that, there is now a fast mode for Opus 4.8 where the model can work 2.5 times the speed and is now 3 times cheaper than the previous models. So the model 4.8 is more powerful than the other models, it can work 2.5 times as quick and 3 times cheaper, which if you ask me is a pretty good deal. So just to quickly run you through Opus 4.8 because that is how we're going to be using this new workflows feature, these are the benchmarks. And of course benchmarks don't mean everything, but they give you a pretty good idea of the jump that it's made from Opus 4.7. It pretty much beats every major LLM now in everything apart from agentic terminal coding, where GPT-5.5 just gets it. But on everything else, including agentic coding, it wins by quite a long shot compared to these other models. But of course, the star of the show today is the dynamic workflow. So I'm going to take you over to this article here posted by Anthropic. It's basically going to run you down what it actually is. So, introducing dynamic workflows in Claude Code, helping Claude take on the most challenging tasks end-to-end. Work you'd normally plan in quarters now finished in days. Claude can dynamically write orchestrations script that run tens to hundreds of parallel subagents in a single session, checking its work before anything reaches you. Some problems are too big for one pass by a single agent, especially in complex legacy codebases. A bug hunt across an entire service, a migration that touches hundreds of files, a plan you want to stress test from every angle before you commit to it. Dynamic workflows can handle all of these end to end. So here it gives you a little preview on what these dynamic workflows actually look like. So a lot of it is done behind the scenes, but it does give you this nice little visual, when it's working with basically all the different phases it's going through. And then here you can see the different JSON files and TypeScript files and all of these agents working inside them. So the way that this is different, I'm gonna get onto this in a second, but the way this is different from just subagents is subagents is Claude code literally just controlling and prompting different subagents. But with dynamic workflows, Claude will actually write scripts that can trigger subagents to work completely independently. So to give you a few examples of, uh, dynamic workflows and actions, it can be done for things like codebase-wide bug hunts. Profiler-guided optimization audits and security audits, large migrations and modernization efforts, critical work you needed checked twice. And by the way, if any of these, especially these first ones, sound a little bit too techie or developer-y, don't worry. The one that we're going to run through in this video, we'll do an example one, it's not going to be anything like that. It's going to be one that you will actually use, uh, almost day to day. So the way it actually works is Claude plans dynamically based on your prompt, breaks into subtasks, and fans the work out across sub-agents running in parallel. The results are checked before they're folded in, and you come back to a single coordinated answer. Agents address the problem from independent angles. Other agents try to refute what they found, and the run keeps iterating until answers converge, which is how a workflow reaches results a single pass can't. So basically, it's just like loads of subagents working on this at the same time. And then before an answer even gets to you, uh, Claude will make sure that basically all of these answers converge and everything makes sense. And there's a whole review process, so you just get the actual correct answer, what you're actually looking for. So they're built for parallel and long-running work that can extend into hours and days. Doing the most complex engineering work that previously would have taken weeks. So obviously you're not going to be using this for like one-shot quick tasks, but for like longer research tasks or like moving over an entire code base or like big, big tasks like this that would, like I said previously, have taken days or weeks, you can now do in a few hours because you can basically just create this dynamic workflow or get Claude to create this dynamic workflow and just let it run and it'll come back to you when it's done. So a nice way to actually visualize this is agent teams look something like this on the left here where we have Claudes talking to Claude talking to Claude talking to Claude. And there's basically obviously in this one, there's three, there could have been more, but essentially it's just multiple versions of Claude all talking to each other and coming to an answer, which of course does still work. But with dynamic workflows, it looks more something like this. We have Claude at the top as basically like the main orchestrator, and then it assigns different tasks. So the task that you're trying to do, it splits that up into subtasks. So let's say there's three subtasks here, it then passes that to a new Claude subagent that's going to be the implementer, and then that Claude subagent will have multiple subagents underneath that that are going to be all of the verifiers. So this is going to be the implementers doing a lot of the work,, and then there's going to be verifiers down here, and there's also going to be more subagents for each of them that are going to be acting as the fixers. And of course, this could change workflow to workflow, but it gives you a pretty good idea of how it looks side by side with agent teams and what actually makes it different. And also, one of the other things is all of these are going to be contained in scripts as well. So Claude will just write a script to be able to spin this all up. It's not to spin literally every single subagent up on its own. Okay, so I'm over in Claude Code now, and there's a couple ways we can actually use this. The first one is if we go to effort And then now you'll see this new— so we've got low, medium, high, extra high, max. And then now we also have Ultra Code, which looks pretty awesome when you hover over it. But basically what that does is it's extra high. So it's going to be like thinking really, really hard. So like the highest level of thinking. And then also workflows are going to be actively kind of like top of mind for it. So it's not like it's going to execute a workflow every single time you prompt it, but workflows are going to be more top of mind for it. So it's going to be more likely to use them because chances are if you're using the Ultra Code extra high thinking mode, you're going to be doing some pretty big tasks. But for now, we're just going to leave it on high. And the other way you can do it is if you just type in workflow here, you can see it gets this kind of like rainbow effect. And that just means this is basically putting an extra prompt in with Claude to say we want to create a workflow. So it doesn't by default trigger. If I just said to Claude, what is a workflow? Claude will be able to say to me something along the lines of, oh, I see that you don't actually want to trigger a workflow. You just want to know what a workflow is. So it's not like blanket, it will trigger a workflow. It's basically just telling Claude that you want to trigger a workflow. So for this, we're gonna be doing a competitor research. So I wanna research all of my competitors on YouTube, see exactly what they're doing. Now, obviously I do this anyway, but I want to get a much, much more in-depth look at what they're actually doing, and I want to have different agents actually assigned to each of my competitors and different agent teams assigned to each of my competitors. So that way it's gonna be way more comprehensive. So I said, can you build a workflow to scrape and analyze all of my competitors on YouTube? I wanna see exactly what they're doing. And how that differs from what I'm doing on my channel and the exact steps I could take to implement what's going well on their channels onto my own. So what we're going to do is we're going to send this off and Claude is going to pick up that this is a workflow. So it's going to start initializing this workflow sequence and actually building out a workflow. Okay, so it said I'll build this as a multi-agent workflow that pulls real data from each channel, runs a gap analysis against yours, and produces an exact implementation plan. Two things that change how I scope the run. So I already have some competitors here that I just have like in my context files inside of Claude Code. So it's basically 'Who do we actually want to scrape and analyze?' So it looks like I already have 3 in here: Ed Hill, Simon Scrapes, and Nick Sarev. So we could do those, or we could do the track 3 plus more. So people like Liam Otley, Helena Lu, and then Brett Malinowski. Apologies if I, uh, pronounce those names wrong, by the way. Or we could do a completely different set, or we could just type something out. So I'm going to say we're going to do the track 3 plus a couple more. So we'll see how many it actually comes out with. And then it's asking us for the output format. So a markdown action plan, a branded video presentation, or both. So I'm going to say it would be pretty cool to have a branded video presentation. So we're going to go with that and then review your answers. We're just going to hit submit there. Okay, now it's just asking us which competitors. So let's say, yeah, Liam Motley, uh, Brett, Helena, and then Jono Catliffe. I'm not too sure, I haven't heard of him before, so we're just going to stick with the ones that we have. And then we're going to go with my YouTube channel here and then submit answers. Awesome. All right, so it said locked in, here's what I'm running. So it's got these 6 competitors here and then my channel and a branded Autonomously AI video presentation that I can actually record. So something I could use in a video. And by Autonomously, uh, it's just my agency, so that's why I keep saying that. The workflow is going to scrape each channel for real data, so things like subs, candence, top video view counts, title/thumbnail formulas, content pillars, and then hooks. Find the cross-channel winning patterns, run a gap analysis against my channel, and then build an exact implementation plan. Then Claude is going to render it into a branded deck. Okay, so now it's actually like initiated this workflow. You can see down here it pops up like this, so we know when we have a workflow running. And it's saying here that it's now running. So the workflow is running in the background. It's just giving us a breakdown of what it's doing again. So workflow is running in the background. Here's what's happening while you wait. Basically, again, just a breakdown of what it's doing and then what the output is going to be, and then when it lands, how that's actually going to render. So now it's basically just going to sit and wait for one dynamic workflow to finish. But if we actually want to see it in a little bit more detail, we can run /workflow, and here you can see what's actually going on. So we can flick between the different phases. Yeah, obviously these ones haven't been started. But this scrape one is starting now. And each one of these is going to be using Opus 4.8 to do the actual scraping. It's not always going to use Opus 4.8 for all the models because for smaller tasks it might use like the Haiku models. But something to note just with this is that workflows do use a lot, lot more tokens, as you probably imagine, because you're going to have like dozens of different agents, um, spun up to do tasks like this. So just be wary of that. And of course, only need it when you actually need to. But for something like this,. It's a really great example of when it'll probably just be easier to just scrape it one time using the workflow, use a bunch more tokens, and that way you don't have to do like more little scrapes and keep going back and forth and they're not giving you the right outputs. And also something to note as well is it's doing all of these in parallel. So you can see this one's idle, which pretty much just means it's waiting for the other ones, but it's doing all of these in parallel at the same time. So it's not gonna wait for that Ed Hill one to stop and then start Simon Scrapes and then do Nick Sarav and do it that way. It's doing all of them at the exact same time. So in theory, they could all finish at the exact same time, but of course different size channels, different type of scraping for each of them. So as you can see, like I said, that third one that was idle has now finished. So Nick Sarav has finished, we just got that little tick there. We'll just wait for these other ones to finish, and it's also going to scrape mine at the bottom, which is quite cool. So once these are finished— another one's finished there, Liam Oatley one's finished there as well. So now about half of them is finished. When they've all finished, it's going to move on to the next phase. Okay, so it's almost finished now. You can see each of them have used about 50,000 to 60,000 tokens. This one was the most, 67,000 tokens. They all used around 15 to 20 tools as well, and they all took around, uh, 2 and a half to kind of 3 minutes. Okay, all of these are now done. So now if we move on to the next phase, which is going to be patterns, and it's going to be using Opus 4.8. So only one agent here, um, so it should be a little bit quicker than the other ones, but of course we'll see because it's going to be having to go through a lot of data. Okay, so it's just finished the pattern recognition as well. So now it's got some patterns. Now it's going to do a gap analysis., and then after that it's going to plan. So for all of these, it's using Opus 4.8. I did a test earlier and it used, um, Haiku 5— 4.5, I believe. I don't use Haiku too much, so I think it was one of those models, Haiku 4.5. And of course it did stuff a lot quicker and it was a lot cheaper. But in this situation, Claude thought that Opus 4.8 would be the best model to do all of these different tasks. Okay, gap analysis done. That was pretty quick. That only took like 10, 20 seconds. Um, and now the plan. So now it's actually going to do a plan for all of this, and it should give us a nice pretty report, uh, that we can have a look through. Okay, cool. So it's now finished that. So if we exit out of this plan view, it should actually give us, uh, the report or however it's going to present it to us. So you can see that, uh, workflow has now disappeared because the workflow is done. Okay, cool. So it's just finished here and it says this is what we've got. So we're going to open this up in a sec. It's basically broken down what is actually in there. So it looks like it's about 7 pages. The headline finding: you're not on the wrong topic. It's the single best performing pillar of 2026. 'Your problem is packaging and candance.' And then it's basically giving me this breakdown, but we're going to open this up and we're going to see what is actually in there. Okay, here we go. 'You picked the right wave, your packaging is sinking it.' So 7 channels analyzed, 2.2 million combined subs, and then my subscribers here, and then my share of the room out of everyone that we, um, scraped. So I am 0.3% of the sub count of everyone that we scraped. So you are a rounding error for now, is saying. 'Right wave, wrong rapper. You pivoted to Claude Code early but neutral titles and one upload a week mean fresh videos land.' 200 to 1,000 views. The good news, every channel here, it started exactly where you are. So roasted me there a little bit, but need to hear it. And then the next one is Simon Scrapes with a 4% share. So rode the exact same pivot you're on. And then of course you can read that breakdown as well. I'm not going to read through all of them because we'll be here forever. And then Helena Lu— apologies if I pronounce that wrong— sells the Hands Off Dreams. She's 12.4%. And then Nick Sarav is 19.7%. And then Brett is 23%. And then yeah, Liam Motley, 36%, as you could imagine. 6 channels, 1 operating system. So packaging How they title a frame: dollar anchors, not percentages. So they stick to dollar amounts in their titles. Contrarian stop commands, so stop doing X, stop making ugly websites, etc. Every X explained. Cool. Beginner qualifiers, so no code for normal people, etc. And then the distribution, so how they actually get really— Newsjack, every drop, a reaction within 48 hours, uh, Shorts as top of funnel, high relentless cadence. And then own a category term, so AI automation agency, Agentic OS. Things like that. Titles and framing— apparently I already own the hard part. So assets I've built, I'm on the right wave, I have a real business proof series, so I actually have a real business doing this stuff. And then I also have a built-in funnel. And then proof it works— you can make Shorts explode. I had one Short that did pretty well, got like 93,000 views. I have a brandable category asset, and the tooling already exists. So basically I already have the resources to make all this good content. So same content, new wrapper. So this week retitle with a dollar anchor. So for example, the Claude Code setup that saves me $4K a month. Ship 1 Newsjack, post 3 Shorts, and then add 1 contrary title. So stop building NAN agents, do this in Claude Code instead. And then in the next 30 days, flagship authority video. So every Claude Code business automation explained for non-coders was an example of one of Simon's one. And then do a free full-length course. So a lot of people are doing like 4-hour courses on there that do pretty well. And then brand a category, so something like Claude Business OS. And then post at least 2 3 times per week. And then repackaging and reacting. So the Claude Code setup that saves me $4K a month in sales admin, copy it. So modeled from Liam Motley's. And then basically all these are all the videos that we're going to be copying, but just obviously changing it around a little bit. And then building the flywheel here to kind of just like repurposing the content that we do that actually does well. Awesome. So that full report there came from that one workflow that really didn't take too long to actually do, but you can see how structured and how in-depth and how comprehensive this actually is. So that's it for this video, guys. Thank you for watching till the end, and I'd love to hear what you guys are actually building with Claude Code workflows.