It’s Sunday. The time is 7:30 p.m., and it is the end of July 2026. I’m sat in my hot tub, and I’m just thinking about the week gone by and all of the releases, as well as just receiving a text from a friend of mine who’s really struggling to get the best out of AI.
On one side of the world, I see people using voice now to control their entire computer, entering a new dawn of computing where we’re stripped away from those mundane old ways of working, such as the keyboard and mouse, and into a new era of what I believe is definitely going to be commonplace with everything that is voice.
And then, on the other side of the coin, I have a friend—I won’t mention any names—who’s struggling to set up AI to actually do anything for her. I noticed the gap between those who understand and those who are really struggling.
I think it comes down to understanding tooling.
Three Groups of People
I think most people see ChatGPT or Claude in three different ways.
The first group of people use ChatGPT or Claude as an alternative answer engine. In other words, they are planning a trip somewhere, or they want to find some information about somebody, and they ask the AI. They pull up ChatGPT, ask it a question, and have it find things to do for them.
The second group of people are using AI with some connectors. They’ve figured out how to maybe add Asana, Salesforce, or Microsoft 365 connectors to their Claude environment, and maybe they’re asking it simple questions like, “Look through my schedule and tell me what my next meetings are,” or, “Help me write this next email for this brief,” or whatever’s next—typical things that we see on a daily basis. Again, low-level answering systems.
And then there is the third group of people, like the people on X who are deep in the open-source world, trying out different harnesses, trying out different APIs, and connecting to software that is specifically for creating videos or 3D models—Unreal Engine’s MCP to create games. They’re using it to create insane web designs that look out of this world, with motion effects and animations. They’re also the people who are using the advanced features of all of the AI products.
I definitely think it is completely unfair for everybody unless you are in group three. Something is only as good as it is if you know how to use it.
Think of a hammer. A hammer can be used to build the wooden foundations of a house, but it is merely just a hammer. Most people are using the hammer to hit a nail into a wall to hang up a photo, whereas the true utility of a hammer is its capability of building an entire house or a home for a family.
So, how do you go from level zero to level three, or from the first group to the third group, as quickly as possible?
I get this a lot because I get asked, “You’re using all this AI stuff. I can’t seem to get it to do anything for me. What are you doing? What am I doing wrong?”
I think it just comes down to a knowledge gap, and the knowledge gap can be filled by understanding these few things.
The Harness
The first thing is the harness.
Claude and ChatGPT are great by themselves. Don’t get me wrong: their desktop applications are useful, most specifically Codex and the latest version of ChatGPT, since you have access to computer use on Macintosh. Again, Windows is really struggling behind here.
Otherwise, you’re very limited to adding connectors in the UI that you get access to, which are usually only MCPs with read-only access, and maybe some prompting with chats and projects, and maybe some skills if you can get that far. You’re very limited in what you can do.
You can’t really do outside work, taking advantage of tools that have been made AI-native and specifically used for large language models to do a lot of the busy work that you would otherwise use a computer for, such as opening a Word document and writing all the information in a Word document. That’s the basic layman stuff of ChatGPT.
If you get the model outside of the harness into something such as Hermes Agent or OpenClaw, you can then start calling on other tools through APIs to do more advanced things, from, like I said earlier, video editing all the way through to 3D renders and any creative purpose.
We know how much Apple dominates the iPhone world. Therefore, I would recommend getting yourself a Mac Mini so that you can pair it with all of your Apple technology that you already have. Those of you who are Windows users, good for you. I know we’re still alive out there, still going strong.
And yes, I have been a Windows user over a Mac user for the last 20 years. I also am an Android user. Yes, I’m a nerd, blah, blah, blah. Deal with it.
Once you get your Hermes Agent installed—and here is the documentation to go and install Hermes Agent, because I really think that OpenClaw has become so bloated now that Hermes has superseded it and passed it—that leads us on to the next thing.
Setting Up Your Tools: Moving From a Thinker to a Doer
If AI’s purpose is to outsource and take control of the work that you need to do, then you need to give it all the skills and all the capabilities possible to be able to do the work. That includes access to the software that does the work for you.
Since computer use has been rolled out, it can definitely do a lot more, but it still uses the old way of doing things, such as OCR: taking screenshots of the software as it goes on a Chrome browser, deciphering what it should select, clicking the next thing, deciphering what it should connect, clicking the next thing, deciding what it should press, clicking the next thing, and so on.
It’s a very slow process for getting something done. The other way, and the best way, is to connect to the backend of the software through APIs. The easiest way to do this is to find the software that you would use. Please note, there is also much better AI-native software built for particular tasks, such as HeyGen and Hyperframes rather than After Effects, to create and edit—to do the busy work that you want it to do.
The easiest way to do this is to generate a bearer token or a personal access token that you can give to your AI in order to connect to that software, navigate through the backend via the API, and take the actions that you want it to take.
A good example would be connecting to Nano Banana’s image-generation tool through Google AI Studio—or, even better than that, Higgsfield, because you have access to all of the image-generation models and video-generation models, including some that aren’t even available on your computer and won’t run on your computer because you don’t have the VRAM or a professional GPU.
You can use API tokens to connect that software and then send the prompt to that software through your agent. Let’s take an example of how that might look in a working day.
First, you might have your AI run a cron job, which is just the name of the tooling given to what most of us know as automations. What happens is it wakes up and decides to do something based on the prompt. The prompt may be something like this:
Connect to my Asana task list. Look through all of the tasks that I have for today and the next few days, organize them in terms of priority, and send me a summary of all of the tasks. Then begin to perform the tasks with all the relevant tools that you have access to.
>
Once you’ve finished these tasks, send me a summary of each task that you finished and update the Asana task comments with all the information about the work you’ve done. Write it from my point of view, and keep it simple and in layman’s terms so people can understand.
>
Tag anybody you need to in Asana to tell them that the work is done and/or close the ticket if it doesn’t need a human in the loop.
For any creative work or public-facing work that is out to the market, I recommend having a human in the loop. For anything internal that just solves the problem rather than creates something, often you don’t need a human in the loop, but I’d recommend doing it for the first few months, until your agent really understands what’s good and what’s not.
That might be a really good cron job that you set up every morning at 5 a.m., so that when you get to work at 8:30 a.m., you can see all of the work that needs to be done today that still needs some kind of human element—which shouldn’t be very much if you get really good at the tooling.
I work for a financial company, and in my role as an AI automation specialist, I work closely with integrations. We use typical, run-of-the-mill CRM software like HubSpot and Salesforce. A lot of the management and workflows that need to be created in these tools can easily be done by an agent and sent back to you as a table to review, with a working link ready to go.
Perhaps there is a campaign that needs to go out to the market, and you need to do market research on that target market. From that target market, you have a specific avatar. Then you take the information from that avatar and have it run through your copywriting framework.
Now you have a good understanding of the target market and the pains and problems that market has. You might even get to understand the language that the target market uses, so you can really resonate with them through your copy. From that copy and market research, you can now start to build your creatives.
These might be ads for native or display advertising or media buying. These might be landing pages to drive these advertisements to, solving these specific pains and problems. Please note, there are different ways to paint their pains, problems, and solutions, and often, by tweaking the headlines around a particular target market’s pains and problems, you can increase conversion rate, sometimes up to a thousand percent.
Having an agent figure out five, six, seven, eight, eleven, fifty—maybe just three, four, seven, or ten—landing pages with different headlines but similar copy might help you figure out how easy it is to convert somebody and the best language to use.
Then you might have some Calendly integration software or something to set up an appointment. Perhaps it might go through into a dialer. Perhaps you might have a workflow on submission. There may be some backend AI automation that you need to do for the client to follow through a process: perhaps just notifying a sales team member, perhaps updating lead economics or lead status, perhaps enriching the lead by sending it through applications like Apify or other AI enrichment tools.
And then finally, after the sale, making sure the customer has a wonderful, world-class experience: delivering the product or service to them in an efficient way, improving systems or operations in order to do that, and ultimately following up and keeping that relationship going, so it’s not just transactional, but actually achieves a client for life.
We’re in the game of clients, not customers. I don’t care what your business is.
Okay, so those are the two things you really need to do to get going with your AI agent. Look at the things you have to do on a daily basis, the tasks that you’re oriented with, the systems that you need to set up, and think about the tooling that’s specifically related to those workflows.
Once you know that, and you’ve connected them to your AI agent through either OAuth or by sending bearer tokens, you can do a couple of tests to figure out if the agent really, truly understands the task. Once you have a completed version of that workflow, make sure you log it and ask it to create a skill from that particular workflow so that it knows how to repeat it. That can be done multiple times.
I’ve used a system like this to constantly check our target market’s keyword research for search terms—be it long-tail terms that our target market is searching for—and then automatically create blog content, spin up a unique blog article, add that article, make sure it’s SEO-optimized for the search engines, and publish it automatically, while generating a nice image from Nano Banana or an image-generation tool that doesn’t feel too AI, in order to increase our SEO ranking for local search results.
Using this technique, we’ve also been able to build new landing pages for salons. We’ve expanded from one to three in just the space of six months because we’ve been able to drum up business by using AI to build entire business startup flows: from doing market research on a target area and making sure that there’s enough clientele for us, to researching the build of the commercial buildings and the real estate that’s available, to sending messages to these Realtors and making sure that all of the problems that appear to be the case in the images and floor plans are not the case—such as room sizes, getting floor plans, understanding municipalities, and understanding the condition of roofs, water tanks, or AC.
Anything that might take human intuition to work through and go back and forth, to offering on the property, to creating landing pages, running adverts, and creating adverts to make sure that a target audience is interested in the business that we have to offer.
All pre-sold, all started from one or two prompts that take this thing into an entire workflow.
If you want to learn more about that, let me know and I’ll share that in the next few lessons. But the point of this blog post is this: your AI tool is only as good as you program it to be. The difference between those who get the most out of AI and those who don’t just comes down to the sophistication of the connections to the tools and the skills that those agents follow when using those tools.
If you want to learn more about how to set up your own AI agent system for your business, let me know. I’ll be happy to help.
I have limited availability, and it’s based on a first-come, first-served basis.
Until next week, have a good evening and have a nice week. Hopefully no new frontier models are launched, and we can still continue to keep up.