I Built My First AI Agent Team — Here's How

Transcript

G’day. My name’s Mike and you’re listening to Lone Wolf Unleashed, the podcast that helps you take time back from your business so you can live it on your terms. This week I have implemented my first agent team to start doing stuff in my business.

I’m super excited to walk you through how I’ve done that and how you can do that for yourself as well. So the system I’ve used to do this is in Paperclip. It is an open source piece of software.

You can go to paperclip.ing I believe it is, to go and check that out. I have initially installed it on some local servers just on my machine to play around with.

And then once I was happy with the fact that I was going to be using this more often, I have deployed it onto my own cloud server, which is pretty cool. And I’ve basically gone and created an agent team to help with some of my content management things.

Things like generating of titles and descriptions for clips and drafting of newsletter articles.

Based on the transcripts of these podcasts, I’ve connected them up to my different tools such as SharePoint and Metricool to be able to do that automatically. So how do we start? Well, I have a target operating model and I modeled this out in Obsidian.

It’s similar to if you go check out my 5PS framework, the second P is profile. I’ve done previous episodes on that. You can check out the 5PS framework on my website.

The profile basically paints a high level picture of your business, lists out a lot of the high level processes, the different platforms and things that you’re using.

This time around, I’ve done a little bit more of a technical diagram that actually connects up all the processes and systems to each other so we can get a sense of how things flow. And at the bottom of all those things is Paperclip. So things will feed into there so the processes can run with the agent teams.

I basically took that, gave that to Claude to say, hey, this is my target operating model. Here’s what Paperclip is and what it does. Give me an initial architecture about how this will work, what each agent will do, et cetera.

So that’s what we did. We took that architecture, I said, great, let’s do up a prompt for the CEO. So the CEO is an agent.

It’s the first agent that exists in your Paperclip instance and the CEO is going to now take that prompt and it’s going to start to form up the different teams with the different capabilities that can exist in there. So the prompt is basically all the work stream items to have that as a project.

So in Paperclip, I went in, I created a new project which is the target operating model. I gave the issue to the CEO to say, hey, this is what we have in mind. What happened?

Well, it spun up a CTO role, so a Chief Technology Officer to help with the understanding of the technical side of things. There’s a business analyst agent in there to help with some of the requirements work.

And then from there it went okay, based on the requirements that we’re seeing here, we’re going to need a content writer agent, we’re going to need a marketing manager agent. We’re going to need a clips publisher agent.

When you look at the org chart, because there is an org chart now of the agents and how they are linked to each other. You can see now the hierarchy in the org chart that that structure has come into, which is really cool.

So then each agent can now be configured to do different things based on the ecosystem that we’re trying to set up. In this case, content management. So it all begins with the content plan. I need to get better at content planning, but that is all done in my Asana.

And so there’s a connection to the Asana, right? So we can draw down whatever the content pipeline is that we have planned.

And then there’s the podcast that we record, just as I’m doing now, that’s done in Descript. Now, Descript has a bit of a limited API in terms of what we can do there.

But once I’ve finished with the recording of the podcast, I can then go and put the recording and the transcript in a specific SharePoint folder. And the agent will pick up that transcript and it will start to do stuff with it, such as understand what the SEO profile of the transcript will be.

It will start to form up various different titles and descriptions of clips and that sort of stuff.

So basically it has helped us understand the types of clips that we can start to produce, where in the transcript, where in the original video we can go to start to take that content out, to be able to snip it up and to get it out on social platforms.

After the clip video is uploaded to a certain folder, another agent will come, it’ll pick it up and it will push it into Metricool by the API and it will schedule a draft post, which is amazing because scheduling some of these posts can be very time consuming and to be able to see how it’s working and sitting there in the background working is very, very cool. What we can do is we can then refine the different configurations. Now, you’ve heard me talk before about how to do Claude skills.

This is very, very much the same. So each agent can be connected to a different large language model or an AI tool.

In this case, I’m in the Anthropic suite already, so I’ve connected it to the various different Claude models. The CEO has Opus, for example. My content writer has Sonnet. One of my agents is even able to operate off Haiku, which is great. Why is that important?

It’s important because, as I’ve discussed in a previous episode, is being able to manage our token costs. The cheaper models, if we’re able to execute the task on a cheaper model, then we should absolutely be doing that. We don’t want to drive a Ferrari to get the groceries.

We want to use the right tool for the right job. So that’s all set up now so we have the different skills, and then we can set this up on a routine. So a routine is every so often.

So at a time interval that we’ve set in, you can go and get the agent to then check the folder to see if there’s a new transcript, and if there is, then do something with it. If not, then it will stop and then it will start again at the next interval. You don’t have to initiate a particular workflow run.

You can just let it pull away in the background and then do the stuff when the particular signal appears, which is amazing. So now we’ve got the process, the overarching process down, the main activities that need to happen, what tools need to be connected.

The thing that I learned about this is because AI is moving at such a speed now in terms of how it operates, what happened was it then had a list of things, a list of issues that came back to me. And issues in Paperclip are just tasks to review, to check, to update different pieces of information.

What I found was it came back with a whole bunch of information that I needed, and it took me a long time to go through.

And I had to engage with Claude about where to get certain pieces of information from for my systems and how to get the API credentials and where to put them securely and all that sort of stuff that took me quite a long time. So it didn’t take the agent teams a whole lot of time to set themselves up and to have the initial stuff there.

What took the time was me then going away and doing the task, doing the review, making sure that things were right, making some adjustments, that sort of stuff. Now, the cool thing though about this is it allowed me and gave me confidence that I was in control.

And I think there’s a lot of people out there at the moment who are scared about what AI can do. We’ve heard the stories about OpenClaw and things about AI that is uncontrolled and it can go out and do stuff. It will just delete file directories.

It will do that automatically of its own volition. We don’t want that. What my Paperclip instance is allowing me to do is it’s allowing me to be in the loop with what the agents are doing.

It has a full run history of what every agent has done at every single point. It gives the thought processes, it gives what it’s doing, what it’s executing, what it’s accessing. All of that information is in there.

So it really does give me a higher confidence in using this moving forward because I am being held and retained within the loop. So that is something to think about as you’re using these AI tools. How much do you know is actually happening in the background?

How much control and influence can you have over the AI as you go about doing things? And I have talked before about how you go about prompting.

Definitely want to be prompting well, but when it comes to doing things that are more routine workflow, this is really important as well, that you’re targeting a very specific folder. Don’t go outside that folder. You don’t need to access anything else. It’s just that folder.

And to be able to see in the order history that that is what has happened is that it’s gone and checked that very specific folder. Okay, found nothing. I’m going to go to sleep again. An hour later, I’m going to wake up, I’m going to check the folder.

Oh, I found something in that folder. Now I know what to do with it. I’m going to initiate these particular skills and I’m going to go and do that.

I hope this has painted a little bit of a picture for you, because being able to document out what our processes are, the types of knowledge that go into executing each one of these tasks, and being able to go in with a solid architecture in terms of how the system will look and how it will behave is super important. You can’t just go from, I have an idea and I’m just going to get AI to set it up. You can’t just do that.

You’ve got to apply your intelligence of your business to the page first. Take the time to sit down and think through what it is that I’m doing, how do I do it, what systems am I using, and then go from there.

AI can absolutely help you brainstorm, it can absolutely help you plan, it can absolutely do all those things.

I’ve got a video that I’m going to be putting on my website that sort of takes you through a little bit of how I’ve set this particular instance up. I initiate a new build in there of a new agent team that I’m setting up around business analysis.

It’s not as straightforward as going, oh well, I’d really love it to just do this and then it go away. You are the master of your business.

It’s likely that if you’re listening to this podcast, you are the sole operator of your business, the sole person in it. You are the master. You know what is going on in your business.

So you need to be able to articulate what is happening so that it then can be replicated by an agent team. Don’t worry about going straight automation end to end. Right now.

Let’s just focus on the little information transfer things that can be done and automated now in phase one, and then we can look at refining later. If you try to go for perfection first, it will take a whole lot longer. So much more effort. Think of the 80/20 rule here.

Okay, we want to get 80% of the results with 20% of the effort. What’s the 20% of effort that I can put in now that can get me almost there?

And then we can make refinements later on in terms of how you automate that. So what have we learned?

We’ve learned document out a target operating model that includes who is doing what in what process and what tech you’re using. A target operating model has those three components. It has the people, the process and the technology.

And then we can give it to AI to do the architecture for you. Give it some of the Paperclip docs so it knows how to behave and what the system does.

And then we can feed an initial prompt into the CEO so it can start to set up your different other sub agents that will need to operate that particular process. Go and check out my website, lonewolfunleashed.com resources. There’s a whole bunch of stuff there.

There’s a whole library there of things that you can go and check out that might be useful for you in building out your business systems. Learn how you can set up your first AI agent suite. It’s a very exciting time in terms of the individual productivity that we are able to achieve now.

Thank you so much, and I’ll see you next week.