The Simple Path to Building Your First AI Agent

Hey everyone, welcome back to another episode of The Human Side of AI, and I'm Flora, your host. Although today is Monday, which means the AI hosts are running the show as usual, I feed them this week's source material and let them take it from there. Today's source is the August 10th edition of the AI with Jam newsletter. If you haven't had a chance to read it yet, consider this your snapshot or a refresh if you already have. This topic is one a lot of you are wrestling with, that anxious feeling that everyone else already has a fleet of AI agents running their business while you're still figuring out where to start. Turns out an AI agent isn't the sci-fi robot or the expensive software suite it sounds like. It's just a file, and building one looks a lot more like hiring and training an intern than writing code. So stick around—no code, real talk, and a story about two AI agents catching a mistake their human boss made while she was in the kitchen feeling refilling her drink. Now let's get into it.

What if I told you the highly advanced, hyper-autonomous AI agent you think you need to buy for like thousands of dollars is actually just a text document sitting on your desktop?

Yeah, I mean, that sounds completely absurd, right? But it's true.

Because look, if you spend any time at all looking at business news or, you know, your tech feeds or even just scrolling your LinkedIn timeline lately, there's this overwhelming pressure. It's almost suffocating.

Oh, totally. The AI anxiety is very real right now. Right.

It's this creeping feeling that you're somehow falling behind in this great AI race. Yeah. I mean, everyone is talking about AI agents.

And they talk about them as if you're, you know, supposed to already have this whole fleet of them running your digital life.

Exactly. Like they should be managing your calendar, negotiating contracts, optimizing your taxes while you sleep. And, uh, if you're listening to this, you might be feeling that exact pressure.

Yeah, I think a lot of people are.

You might be picturing this like hyper-autonomous sci-fi robot or some incredibly complex super expensive enterprise software suite that you have to download and somehow learn how to code.

Well, it is a very real, very prevalent anxiety. And honestly, the terminology itself is partly to blame.

The word agent.

Yeah, the word agent. It implies autonomy, you know, it implies independent thought. And like, most intimidating of all to the average professional, it implies this steep technical learning curve. Oh, 100%. We hear agent and we instantly think of a black box that operates completely outside of our control, written in some programming language we don't understand at all.

Well, today we are going to completely dismantle that anxiety. Uh, we are taking a deep dive into a freshly published, incredibly practical guide from the Real AI with Jam newsletter.

Yeah, specifically their August 10th edition.

It's a great read. It really is. And our mission for this deep dive is clear. We want to completely demystify AI agents, strip away all that tech intimidation, and reveal how you can actually build one without writing a single line of code.

Which is huge for so many people.

Okay, let's unpack this because I mean, when I hear the word agent, I am still naturally picturing that heavy, complex software. But this source makes a shocking claim right out of the gate. It completely redefines the term. It does.

It provides a really vital reset for anyone feeling overwhelmed by all this. The approach starts by just clearing the board and defining exactly what an agent is not. It is not a product you buy off a shelf. It is not an app you download from a digital storefront. And it is certainly not a sentient robot. So what is it then? Well, broken all the way down to its foundational architectural elements, an AI agent is actually just a file.

Wait, just a file? I mean, I have to stop you there.

Yeah, just a file.

A text file cannot negotiate a sales contract or, you know, manage my inbox. You're oversimplifying this, right? What is actually doing the thinking here?

No, that is the crucial distinction. The file itself isn't doing the thinking. The file is the blueprint. It's a document containing your specific instructions, a list of your available tools, and like historical examples of your best work.

So it's just text, right?

The magic, the actual cognitive engine, that only happens because you point a large language model, an LM, at that file to make it come alive and execute those instructions. Oh, I see. The LLM is the engine, but the text file is the steering wheel and the map.

Okay, that immediately lowers the blood pressure. Yeah. Which brings us to why we need a totally different way to think about this. Because if an AI agent isn't a robot and it isn't a piece of boxed software that you just install, we desperately need a new mental model to understand what building one actually looks like in practice. We do. And the source provides this brilliant analogy. It says building an AI agent is exactly like hiring an employee.

And this analogy shifts the entire paradigm. It really does. It takes agent building out of the realm of like computer science and puts it squarely into the realm of basic business management.

Which is so much more approachable.

Exactly. If you have ever hired somebody or even just trained a new team member, you already possess the exact skill set required to build an AI agent. The sequential steps are almost identical.

Let's dig into this specific example used in the newsletter because I think it grounds this abstract concept perfectly. Yeah. The intern example. Right, the intern. So imagine you hire Sarah from down the street. She's an intern. Now Sarah is smart, she's capable, but she does not know a single thing about your business. Nothing at all. She doesn't know who your core customers are. She hasn't memorized your brand voice, and she certainly doesn't know your pricing tiers. You wouldn't just sit Sarah at an empty desk on her first day, point at her, and expect miracles to happen. Right.

If you did, that would be a catastrophic failure of management, not a failure of the intern.

Exactly. Yeah. You have to equip her, First, you give her tools. You hand Sarah a laptop. Basic necessities. Yeah. And with that laptop, she gets an email address, right? She gets the login credentials to your HubSpot or, you know, your CRM. And then— and this is the part everybody seems to forget when they start prompting AI— you give her context. The context is everything because you aren't going to be sitting over Sarah's shoulder every single second of the day. She needs standard operating procedures or SOPs. Like if a customer emails demanding a refund and you were at lunch, how does Sarah know what to do?

She doesn't unless you tell her.

Right. She knows because you wrote the refund policy down for her in a training manual.

And this is the core insight of the whole framework. An AI needs this exact same onboarding process. It needs the manual. It needs to know what tools it has. It needs the access to use them, and it desperately needs those standard operating procedures. That is the entire build phase. So no coding. No coding at all. The process doesn't start with opening up a terminal or like an IDE. It starts with writing a job description.

So we are essentially treating this like an HR task, not an IT project.

Precisely. You start by defining the dream employee. If you want an agent to run your digital marketing, you sit down and write out the exact job description for that role. Okay. You dictate that they're responsible for, say, SEO blog formatting, writing weekly newsletters, restructuring meta ads, ads, and you do this before you ever even log into an AI platform.

You literally write a job posting.

We write it out as if you were posting it on a hiring board. Then you skip the interview phase because obviously you do not need to interview a piece of software to see if it's a culture fit. Yeah. And you move straight into onboarding.

Okay, but wait, if building an agent requires giving it strict rules and all these incredibly detailed SOPs, just like human onboarding. Mm-hmm. How do you actually do that without losing your mind?

What do you mean?

I mean, I have to push back here. Writing SOPs from scratch is notorious for being an absolute nightmare. Oh, it's awful. Yeah. The curse of knowledge means we naturally skip steps when we try to write down things we do intuitively. That's very true. So if the whole goal of having an AI agent is to save time, doesn't spending 3 weeks typing out exactly step by step how I format a blog post completely defeat the purpose?

You'd be entirely right if you were manually typing it out on a blank page. Well, how else do you do it? That is exactly where most businesses abandon the process, honestly. But there is a genius backward trick to solving this in the newsletter. A backward trick? You don't write the SOP at all. You have the AI write it for you by having it interview you about how you already do the job.

Wait, reverse the roles.

So I just sit there and talk? It completely flips the dynamic. You open up an LLM with voice capabilities like the ChatGPT app on your phone, and you tell it, I want you to act as an eager new intern who is taking over my blog formatting process. Oh, wow. Yeah. And then you say, ask me questions one by one until you have enough information to write a perfect step-by-step standard operating procedure.

That is incredibly smart.

Right? You just talk to it in normal conversational sentences while you commute or walk the dog. You explain your process out loud.

And it just figures it out.

The AI asks you clarifying questions like, What font size do you use for H2 headers? Or who do I CC when it's published? You answer them and then you ask the AI to synthesize all of that into a formal SOP document.

That is wild. It's essentially reverse engineering your own brain, but doing it conversationally. Exactly. You're extracting all that undocumented expertise, that institutional knowledge that usually just lives in your head. And at the end of a 20-minute chat, you have a flawless training manual for your new digital employee.

It bypasses the blank page syndrome entirely. But, you know, this raises an important question. It is. Once we have generated all these brilliant, detailed SOPs for all our different business functions, how do we ensure the AI actually retains them long term?

Okay, here's where it gets really interesting, because if you know anything about how LLMs operate today, you know, we immediately hit a major technological roadblock. The memory issue. Yes. The architecture of these models involves something called a context window, which essentially acts as their short-term memory limit. If you want to build an agent for marketing, another for sales, and a third for daily operations, you are going to max out that memory extremely fast. You really are. You can't just dump all these intricate instructions into one long, endless chat thread. The AI will eventually start hallucinating or completely forget the foundational rules you gave it at the very beginning.

Exactly. The AI's short-term memory degrades over the course of a long conversation, which is why the framework introduces an architectural workaround around that crucially still avoids coding. Love that. You bypass the chat window entirely. You pull the memory out of the cloud and store your agents in a local file system on your own computer.

Okay, let's look at how that actually functions. Yeah. Because the author of this newsletter uses a program called Obsidian to do this.

Yes, Obsidian is great for this.

And the key reason they use Obsidian is that it operates using plain Markdown files. Which is so important. For anyone unfamiliar, Markdown is essentially just the purest form of text.

There is no heavy formatting, no hidden code, no messy rich text stuff.

Right. Because it's incredibly lightweight, the AI can read it lightning fast without burning through all your usage limits or token allocations. Exactly.

But the specific software doesn't actually matter though, right?

Notion works, Google Drive works, GitHub works. Yeah.

Anywhere you can store text.

The author actually has a personal digital vault on their computer named Sutton, and inside this vault, Every single agent has its own dedicated folder containing its specific SOPs and tools.

It's an elegant architectural choice because of the stability it provides.

It's like giving your AI a permanent filing cabinet, but you've also given it a master key and a daily schedule so it knows exactly which drawer to open and when, even if you aren't in the room.

That's a perfect way to put it.

It never gets confused about its role because its entire identity is isolated in that folder.

And if we expand on that, this filing cabinet method provides a massive enterprise-level strategic advantage. It's a concept called being platform-proof. Platform-proof. Think about how volatile the AI industry is right now. We are in the middle of the AI wars. Everyone is heavily relying on Claude at the moment, but, you know, ChatGPT might release a new model tomorrow.

Oh, they probably will.

Codex is doing impressive things. Copilot is integrated into Windows. The landscape shifts underneath our feet every single month.

It really does. It's exhausting trying to keep up, and you definitely don't want to build your entire business infrastructure inside one company's proprietary ecosystem. Definitely not. If they change their pricing or their UI, you're trapped. Vendor lock-in is a huge risk right now. Exactly.

But if your entire business logic, your meticulously crafted SOPs, your historical examples, your job descriptions, if all of that lives purely in Markdown folders on your own local hard drive, you are immune immune to that volatility because you have the files.

You own the brain.

You just point whatever model happens to be the best this month at your local folder. So simple. For example, in a platform like Claude, you can simply click the little folder icon, upload the specific agent's directory, and you never have to re-explain your brand voice or what you sell. Wow. The new model reads the folder and is instantly up to speed.

That is massive. Yeah. You aren't held hostage by OpenAI or Anthropic or anyone. Okay, so we have the architecture down, we have the markdown files, we have the rules extracted from your and we have the LLM connected to the local folder.

We've got the foundation.

But we need to translate this into reality. What does this setup actually achieve in the real world when you take your hands off the keyboard?

This is the fun part.

There's a story in the material that perfectly illustrates this, and it honestly blew my mind.

It's an excellent case study in autonomous agent collaboration.

Yes. So the author has multiple agents set up in their local system. Two of them are named Bosma and Karl. Great names. Bosma runs the marketing folders and Karl runs the sales folders. The author types just one single sentence into the system. They prompt a meeting between Bosma and Karl to make sure everything is technically sound and on track for an upcoming Black Friday campaign.

A single sentence prompt.

And then, and this is my absolute favorite detail, the author literally gets up, walks away from the computer, and goes to the kitchen to refill their strawberry drink.

That physical separation is the whole point. It highlights the true utility of the system. Once it's properly configured with those SOPs, the human can finally step away.

Right. They are in the kitchen. Yeah. And while they are gone, these two agents are interacting. They are cross-referencing the newly built campaign files against the established SOPs in their folders.

Having a whole meeting.

Exactly. When the author comes back to their desk with their drink, the agents have flagged a massive catastrophic broken funnel. They report back that the brand new offer page that author just built for Black Friday isn't listed on the website's sitemap, it has no meta tracking pixel installed, and the email capture fields are completely missing.

And, uh, the most vital takeaway from that discovery: all 3 of those errors were completely accurate, and all 3 were entirely the author's fault.

It was pure human error.

Exactly. Let's analyze structurally why this works so well. The author, being a flawed human being, got excited. As we do. They were heads down building this new Black Friday offer, thrilled about getting it done ahead of schedule. And in that emotional excitement, they completely bypassed their own standard operating procedures.

They just skipped the boring stuff.

They skipped the boring administrative steps. But the agents, the agents don't get excited, right? They don't experience fatigue. They don't rush because it's a Friday afternoon. They simply look at the foundational rulebook, the SOPs the author had previously defined in those markdown files, and they compare the current work against it.

It's so cold and calculating, but in the best way.

They caught the broken funnel because they strictly, mechanically followed the rules of what good looks like.

It's the ultimate safety net for our own human flaws. But it doesn't just stop at proofreading and catching mistakes, right? The material highlights incredible scheduling power too.

Oh yeah, the automation side.

It mentions that Bosma, the marketing agent, writes and deploys an approved SEO blog post every single morning at exactly 7:00 AM. Like clockwork. And there's another agent named Vicki whose entire job runs at 8:37 PM every night while the author is literally asleep. I love that. Vicki goes through the day's work and makes sure every other agent's project briefs are properly updated. But okay, I have to pause here. Yeah, because earlier we established these are just text files in a local folder. A text file doesn't have an internal clock, right? It doesn't. It can't wake up at 8:37 PM. How is this actually happening without custom code?

That is the missing link, and it's a great catch. The text file is the brain, but it needs a nervous system to act autonomously.

Okay. A nervous system.

In a modern no-code setup, you bridge this gap using automation platforms, tools like Zapier or Make.com. Oh, Zapier. Yeah. These platforms allow you to connect different apps using simple visual logic blocks.

Ah, okay. So you set the trigger in the automation tool, not the text file.

Precisely. Yeah. You log in to Zapier and set a time-based trigger for 7:00 AM. Simple enough. The next block in the sequence tells Zapier to grab the specific SOP from Bosma's local folder and send it to the Claude API. Got it. Claude executes the instructions, writes the blog post based on the rules, and sends the finished text back to Zapier. The final block in Zapier automatically publishes that text to your WordPress or Webflow site. Wow. You haven't written a single line of code. You've just visually connected the folder to the LLM and the website on a schedule.

It makes so much more sense. It's just an ecosystem of no-code tools talking to each other. Exactly. Hearing about Bosma, Carl, and Vicki running these intricate workflows while you sleep. I mean, if you're listening to this, you are probably incredibly energized. I'm sure they are. You're probably ready to pause the audio, jump into ChatGPT and start building folders.

Oh, hold on though.

Right. Because before you do that, the framework provides a very crucial warning about the absolute first step you need to take. And honestly, it's going to sound completely counterintuitive.

It might be the most valuable piece of advice in the entire analysis.

The homework is this: stop asking what AI can do. Just stop. Stop looking at the latest model release and wondering what it's capable of. The very first step requires you to leave your laptop at home. Leave it home. Go to a coffee shop with just a pen and a piece of paper and map out the actual functional roles in your business that currently need filling. It's so analog. It is. Write down the software tools you currently pay for and more importantly, write down why you have them.

It forces you to audit your actual operational needs rather than chasing technology for technology's sake. Yeah. The material shares a really telling anecdote about this specific coffee shop exercise. The author runs this paper and pen audit with members of their community, and almost every single time someone has a startling realization about their overhead. Oh yeah.

There was a specific community member who looked at their bank statements during this exercise and realized they were paying a monthly subscription for a Mailchimp account that they had literally never even opened. Never even opened it.

Yeah. And that happens because they bought the tool thinking the tool itself was the solution, which is so common. They thought buying Mailchimp would magically create a newsletter strategy without ever defining a system or assigning a role to manage it. You have to get your systems clear on a piece of paper first. Only then do you introduce an LLM to a well-organized file structure.

So what does this all mean for us? When we boil this entire deep dive down, the grand takeaway is that AI agents aren't expensive, mysterious software that you buy. No, they aren't. There are systems that you clearly define. They are your own rules, your own historical examples, and your own tools organized neatly in a lightweight folder with an AI acting as the engine to breathe life into them.

That's the perfect summary.

And by the way, if you want to actually build one of these alongside a community of people doing the exact same thing, it's mentioned in the source that you can join their group inside ClubJam on Skool to work through it together. It's totally accessible to anyone willing to write an SOP.

It really is a profoundly empowering way to look at modern technology. You are not hopelessly behind in the AI race. You just needed the right mental model to approach it. But, you know, if we connect this to the the bigger picture, I think we have to acknowledge a fascinating, almost philosophical tension here.

I like the sound of that. Where are we going with this?

Well, we've just proven unequivocally that you don't need to be a programmer to build highly effective AI agents. Right. You can build these interconnected systems to flawlessly execute your past strategies and your perfectly documented SOPs. Bosma and Carl will follow your rules day in and day out without ever making a typo or missing a tracking pixel. Which sounds like a dream. It does, but it leaves us with a much deeper operational question. Okay, I'm listening. If your marketing, your sales, and your daily operations are eventually run entirely by automated systems that are designed to strictly mimic what good looks like based exclusively on your past examples, how does a company leave room for true messy human innovation?

Wow, that is heavy. I mean, think about it.

If an agent's entire architecture is built to perfectly follow the established rules every single time, who is left in the building to break them and invent the future.

That is a phenomenal point to end on because as much as we all crave flawless execution and, you know, zero broken funnels, the real magic in business often happens when we make a mistake, when we experiment, or when we purposefully color outside the lines of our own SOPs. So for everyone listening, as you head to that coffee shop with your pen and paper to map out your new digital employees, keep that tension in mind. We started this deep dive talking about the intense pressure of the AI race and that paralyzing fear of a hyper-autonomous sci-fi robot coming to take over your job. It's just a file. It turns out the robot is just a text file waiting for your instructions. The real challenge isn't learning how to code. The real challenge is figuring out exactly what you want the machine to do and what creative chaos you need to keep for yourself. Thank you so much for joining us for this deep dive into the architecture of agents. Keep questioning, keep learning, and we will catch you next time.