Manage DOZENS of Agents At Once with Claude Cowork (Full Workflow)

[0:00]Up until recently, these AI sub-agents were really tools for developers, but all that is really changing fast with the launch of Claude Co-work, with the launch of Claude's Opus 4.6. You can now really 10X your results by starting to orchestrate these groups of sub-agents to do some fairly complex things. In the framework of the orchestra, itself has been a helpful way for me to think about how to get the most out of these new tools. You can think about the traditional way we've all been using AI a little bit like a soloist playing an instrument. We are putting in one prompt, making one request, getting that request back, and just kind of moving forward in an isolated, you know, sequence. This guy Rune just dropped something on X that I think aligns perfectly with this. He says, whatever level of abstraction you are handing off to your agents, you should probably be doing one level above that. And that's exactly right. So moving from, you know, maybe just one prompt and one response to maybe just an ensemble of a few different agents doing some things and then thinking about what a whole orchestra might be. Our moving in that direction of one level of abstraction higher and higher and higher. This framework includes four layers, your instruments, your musicians, your conductor, which is you, and the concert hall. So let's dive right in and start talking about the instruments. Your instruments are basically your foundation models, your Claudes, your Chat GPTs, your Geminis. This also includes any specialized tools you use for transcription, you know, dictating directly to your computer. You know, those tools can include search and image generation. But this also includes your text stack. What are the different tools that you're using? You know, what's your CRM? All of those things I would consider as your instruments. I'm jumping right into the cheat sheet. I make a cheat sheet like this for every single video that I make. There are now over 170 of these instantly available to anybody who supports this channel on Patreon. There is a link in the description for all of that. If you're getting something out of these videos, please consider supporting me there and this is the prompt that I want to share with you. I call this the instrument audit prompt. You can pop this into whatever AI that you're using and get a feel for your tech stack. This can kick off a conversation where you can really figure out what instruments do you use and audit these. You know, you probably have them floating around in the back of your head, but before you start prompting and building sub-agents and so forth, I think it makes sense to dive in and really understand, hey, what are the instruments here that I'm working with? You know, what are the foundational models that I like and that I have access to, what are the specialized tools that I use for transcription, search and then what is my SAS stack look like? Now, probably the most important part of the tech stack is those foundational models. So understanding what they're good at on a deep level is really starting to pay some good dividends. It's worth figuring that out. Here's a resource that I came across that highlights what these different tools are good for. You may know a lot of this stuff, but I'm finding for sure that if anything is related to search, I'm using perplexity. You can use an API access. We'll get into some of that more technical stuff here in a second, but anytime I need to search the internet, perplexity is just crushing it compared to these other models that can also search the internet. Even Gemini, I would say perplexity is better at searching the internet. And there are many nuances, so I have a friend that swears that, you know, when he's building software, that Gemini has a much better visual sense and it can see and design, you know, UI and user experience, layouts better than any of the others. And that makes sense because it has such a good model with Nano Banana and some of those others. Claude is just an all-around workhorse. It's definitely my main tool these days. And Cursor itself recently ran a very long experiment into autonomous coding and found that the nuances of these different AI models really made a difference when they built a bunch of different sub-agents that worked for like a month straight. So understanding these tools at a basic level and then, you know, understanding them at a deeper and deeper level is probably worth your time. I know it's tempting to just say, I use Claude for everything, that's kind of what I'm doing these days, but as I'm doing more and more research, I'm starting to learn, okay, let's use Codex for certain things. Let's use Gemini for certain things, let's use perplexity for certain things. So understanding your instruments is really the foundation of this orchestra. And once you understand your instruments, it's time to hire the musicians, and this is probably the most important part. This is the core of what I want to get at today. This is building and organizing these sub-agents. You can think about giving each of these different agents a name, a role, a scope, tool access, and a behavioral guide. It's like writing an employee handbook for each one. The key insight here is that the constraints matter just as much as the capability. Your agent configurations can become your moat over time. The easiest caveman way to spin up these agents is right here inside of Claude Co-work, which is the desktop app that Claude recently released. So just for a quick demonstration here, I'm asking it to please spin up four separate agents to simultaneously research 1 - AI industry news, 2 - news about the Boston Celtics, 3 - Local Maine News, 4 - History on this day that might be interesting to an entrepreneur, and I'll try to spell that right. And we'll spin this up and I will show you that it will begin creating these simultaneously. So start thinking about what processes can you run in parallel. There are so many ways to benefit from this that it's still boggling my mind. Here they are. You can see them right here, four different agents running through multiple steps. Here, we can click in a little deeper and see all the steps that they are running, all the different things that they're looking at, which is really, really cool. So there's a lot of this that is automated right here in Claude Co-work. Sometimes for longer tasks, you will see things pop up over here in the progress bar where you can chime in and help guide these agents as they're moving along. And if you want to get one step fancier, which I think you probably should, think about creating a file structure like the one I have here where you have instructions for your different agents inside specific folders here. So you might have one for any of your, you know, research agents, you can use a file like this, title it claude.md, which might have instructions about how you want your research done, exactly what steps to take and what sort of research you're most interested in. Um, you can do the same thing, obviously for report creation, or process agents. You can name these files, these instruction files, different things. Um, some people call them readme's, some call them contexts. A lot of folks are using just claude.md, but Claude should pick up on that if it's named one of these things, um, and begin to execute your commands based on the instructions in there. You can think of those just like any sort of a big prompt or any sort of instructions that you might have inside of a Claude project. That's just the context of the workflow that you want that agent to follow. You can do all of this stuff inside of Claude Code, and you even get more capabilities in there as well. But you don't get this nice layout. You're going to be stuck looking at this terminal. But it's really not that bad considering all the power that you get from this. So if that's what you're dealing with, find the folder structure that you built here, do a right click on it, and then, uh, on Mac, you hold down option. I'm not sure exactly what you do on Windows. Then you copy the path name and just go CD, paste that path name in, hit enter, and then fire up Claude. So that puts you into that folder that you're working in. And now you can do everything that anybody in Claude Co-work is doing and showing you, plus you get access to a lot of these powerful code tools as well. And this really is the next level of controlling those agents. You can do a backslash in here and do agents and you can then enter into this agent's section here, and with using Claude Code, create a new agent and we're going to do that, uh, at the project level here, or you can do it at a more global personal level, but this is just at the project level in that folder that I'm working in. And then you can begin to generate this and work with this agent, um, build this agent with Claude itself. So that is kind of the highest level of starting to build these agents and starting to get them to really interact deeply with the instruments from the last section that we talked about, integrating with all the various AI models that you know and love, as well as all the other SAS tools that you might use and then the APIs that run behind the scenes of those. There are a lot of good prompts inside of this cheat sheet for designing your different agents and the different ensembles that you might want to put together of these agents based on whatever type of work you do. There's also a really cool website that I found sub-agents.cc, where you can see all of these different agents and the ways people have put them together. Uh, so you can peruse that as well, really cool resource that I found while I was researching this video. So you've got your instruments, you've hired the musicians, and they're warming up. Now it's time to turn that noise into some beautiful music and step into this role as the conductor of the orchestra. You're probably quite familiar with sequencing different workflows from one prompt to another, but when you begin to conduct the orchestra, you're now sequencing waves of operations. The way that I've been telling the AI how to do all this stuff is in these larger workflow documents that describe, you know, wave one, fire up these agents, and then it'll know to go into those folders and find the different instructions in those folders, complete wave one, collate everything from wave one, which again, might be, you know, researching all different areas and then pulling the best of all that research into one sub-agent that can look through all of that and begin to collate that or make different decisions based on that, then a checkpoint. So, uh, adding uh, the user into the loop, saying, hey, from all of this, how do you want us to proceed to this next wave that might then take all that research and turn it into some sort of a documentation, uh, maybe firing up several different sub-agents to work on different parts of that documentation. The sequence is those orders of those waves. The routing is, you know, telling it which agents to give the work to, and then the handoffs are the moments that one thing moves to another. How do you want things to be handed off and when do you want the user to be looped in to make various decisions. Here's a prompt that you can use to work through all of that with your AI agent, just saying, hey, I want to build this orchestrated workflow. Let's design a sequence map, agent assignments, handoff rules, guard rails, failure modes to begin building that document. And I would add that workflow at the root level of your folder structure here, so it can find that, and then it can see into how all these different agents might operate. All right, so the orchestra has been assembled, the conductor has created the score, but where does this performance actually take place? I like to think of this whole folder structure and this knowledge base that we're building as the venue, as the concert hall where all of this stuff is happening. But when you start to work with it, things are going to change. You're going to start to get feedback loops. This is the audience participation where they're clapping and you're getting good results, or they're booing and you're figuring out, okay, there were getting errors and, you know, these responses are not quite what we want. We can then think of those as the feedback loops that can then help us to improve this process as we go along. Maybe we're publishing something out there on social media and it's not getting any likes. Well, let's figure out what that is. Let's loop that in as a feedback loop there and bring that back into our venue and begin to improve this process based on the reflections of the acoustics and the audience response to our work. I've got a cool prompt in here for how to organize your concert hall. Another twist on just how to organize your knowledge base. And then this one is this after-action review. So a lot of the problems you run into are going to be pretty obvious as you're working through stuff. And every time I do a process, I'm realizing, oh, I can make it a little bit better, I can tweak it a little bit better. So now I've come to realize that is the process, continual improvement is how we're going to stay on this steep learning curve of all these tools and different processes that are getting more and more advanced. That's not the exception, that's actually the norm. Here's another cool resource that I found. You can either grab this out of the cheat sheet or Google this on how to build self-improving AI agent through feedback loops, and you can start to think not just of improving those agents individually, but improving your entire workflow and all of the different ways you're orchestrating all these different sub-agents, because now so much can be done in parallel that you want that to be improving over time. Each time you're running it, making a little bit of an improvement and they'll start to compound over time for some serious benefits. One of my favorite AI dudes, Ethan Molick recently talked about how, you know, this skill of orchestration is not a technical skill at all. It's an organizational skill. It's sort of a management skill more than anything. So if you've been a manager, you now have, I think, some advantages over some highly technical coders who don't know how to think that way. Because you've gone through that process of, you know, finding the right roles, getting the right people in the right seats, and motivating them, that's just all those skills are going to transfer into this new world of AI. The only difference is that these workers run on tokens instead of coffee. So dig in there, get started, get familiar with your instruments. Start training your musicians and these different sub-agents, begin building these big workflows and step into that role of the conductor and start looking for those feedback loops that the audience gives you, that the process gives you so you can continuously improve that over time. You can follow along in the cheat sheet. This should really take you by the hand and help you build all this stuff out. It's instantly available again to anybody who joins my Patreon. There's a link in the description. I've also got some really fun group calls that happen on a weekly basis where you can work directly with me, ask me questions. There's really no question that goes un-answered. So jump in there. Let's work together and, uh, I'd like to point you to another video all about Opus 4.6 and some of the other things that it can do that were not possible before. That video's right here. I will see you over there.