00:00:00Now that AI agents are becoming more and more useful and more and more of a thing at least
00:00:05for some tasks, I think it's really interesting to see how we're kind of going full circle.
00:00:11And here's what I mean. If we take a look at the history of computers and the internet
00:00:16as a whole maybe, we could maybe draw a ease of use chart that looks something like this.
00:00:23Of course this is totally made up but you probably get my point here. We started in the 1970s
00:00:30or somewhere around there โ don't nail me on a specific year โ but in the early
00:00:36days when you and me โ well, I wasn't even born โ but when normal households didn't
00:00:41even have a computer, interacting with computers was mostly text-only through terminal user
00:00:47interfaces through the command-line essentially. And rich graphical user interfaces and rich
00:00:54websites and all that fun stuff โ operating systems built for normal users โ that only
00:01:01became a thing around the 90s, 2000s and of course kept on evolving until today. And it
00:01:09still is of course evolving, don't get me wrong, I'm not saying that this is all going
00:01:14away, but one thing that is clearly visible and easy to see is that with AI agents we
00:01:22have this strong trend back to text-only input, terminal user interfaces, CLI tools, markdown,
00:01:31JSON, all that basic stuff. And with that I don't just mean that we have tools like
00:01:37Claude Code that don't really come with a graphical user interface โ though there is
00:01:43a desktop app but it's primarily consumed as a command-line tool โ I don't just mean
00:01:48that. Instead I mean that all these AI agents, these agentic tools โ however you want to
00:01:54call them โ really excel at interacting with other command-line tools, other programs they
00:02:02can invoke via the command-line, they want simple text, simple formatted text like markdown,
00:02:09that is where they really really shine. And that's why more and more companies โ like
00:02:15for example a few hours ago when I'm recording this, Google released more and more command-line
00:02:21tools. Like Google released a Google Workspace CLI. Believe it or not, that didn't exist
00:02:27until now and it's a tool you can use to interact with your Google Workspace services like Gmail,
00:02:35Google Drive, through an official CLI. Now there already were other solutions โ like
00:02:41for example GOG CLI by Peter Steinberger, the creator of OpenClaw โ he built it because
00:02:48he wanted a programmatic CLI-based way of interacting with Google services and that didn't exist
00:02:54until a few hours ago but now it does exist. And this is not a sponsored video by Google
00:02:59or anything like that, it's just interesting to see that more and more companies that offer
00:03:04services are releasing tools like this. MCP servers would kind of be a similar thing,
00:03:11though MCPs in my opinion have various disadvantages and I strongly believe we'll see CLI tools
00:03:18and APIs and in the end CLI tools just wrap APIs in the future or as the main way of consuming
00:03:27services through agentic tools. And here's a concrete example for what I mean. Over the
00:03:32last couple of days and weeks I've been playing around with the PyCoding agent. Now
00:03:37the PyCoding agent is, you could say, an alternative to codex clot code. It's simpler in a good
00:03:46way, more limited regarding its features but very very powerful and you can use it with
00:03:51your codex subscription for example. Now this video is not primarily about this agent and
00:03:57it doesn't matter really if you use this or codex or cursor or whatever, they all can
00:04:01get you there. But I also like this tool a lot and most importantly, just like clot code
00:04:07and codex you can also use this tool for non-coding tasks despite its name. For example, it's
00:04:13actually this Py agent that's being used internally by OpenClaw. So that's the heart,
00:04:19the logical heart of OpenClaw you could say. And then OpenClaw of course added way more
00:04:24to it like memory and channels like Telegram and WhatsApp and all of that fun stuff. But
00:04:30this is one agentic tool you could be running on your system to do stuff. You could also
00:04:35build your own agent of course. And I got a course on that where I also cover how AI
00:04:40agents actually work and what the difference to workflows is and often you maybe want a
00:04:44workflow and not a true agent. But I got a course on that if you want to dive a bit deeper
00:04:49into that. I also got courses on clot code and codex if you want to learn more about these
00:04:54tools. But no matter which tool you're using, what's really really interesting and clear
00:04:58to see is how well they can interact with other CLI tools. Which makes a lot of sense
00:05:03because they've seen plenty of CLI work of using CLI tools like curl, like our command
00:05:10line commands like cd, ls, you know all these Linux commands. They've seen plenty of that
00:05:16in their training data. And they've not just seen standard Linux commands which they therefore
00:05:21know by heart. But most importantly they saw how to use these tools. How to chain CLI tools
00:05:28together. How to pipe results from one tool into another tool. They saw all of that and
00:05:35they excel at that. They also saw that they can use --help to learn more about a tool.
00:05:41And that puts them in a great position of using new tools as well. Tools they haven't seen
00:05:47in their training data like this new Google Workspace CLI for example.
00:05:52Of course if you want to use that through an agent it didn't see that in its training data.
00:05:57It doesn't know how to use that. But if you point it at it, if you maybe give it a link
00:06:01to the official docs, but even if you don't do that it will most likely be able to figure
00:06:05out how to use it by using --help and going from there. Because it's just yet another CLI
00:06:11tool. And large language models in the end excel at understanding and describing and using
00:06:17these CLI tools.
00:06:20And for example the other day, yesterday actually, I had a little problem. I needed to upload
00:06:26a PDF document to a website. And you know these sites that want you to upload a bunch of documents
00:06:32all in one document and that document must not exceed 5 megabytes in size? Yeah, I was
00:06:38on one of those sites. And naturally I had to compress that PDF document.
00:06:43Now I could have tried to find a website that does it for me. I'm not a huge fan of uploading
00:06:49my stuff to some random website though. So yeah, not sure. I could have also checked if
00:06:55there is some tool in my system that can help me with that. But I don't have the Adobe subscription
00:07:01anymore so I would have taken some research. And maybe in the end I would have uploaded
00:07:07it to some shaky website. Well, not with AI. Of course, I could have used Codex or Clot
00:07:13or Cursor or whatever to vibe code a little conversion/compression tool. That probably
00:07:19would have worked as well. But what I did, I span up the PI agent which I'm using with
00:07:26my Codex subscription. And I just ask it to take a look at that PDF file and please compress
00:07:33it while maintaining quality as much as possible.
00:07:36That was all. That was my only prompt here. And it essentially went to work, executed
00:07:41a bunch of commands in the command line, ran some little scripts. And by the way, I'm running
00:07:46this on my system, but I installed a guardrails extension. PI or PI is the actual name has
00:07:53this concept of extensions which you can install. So I installed an extension that prevents the
00:07:59agent from just erasing my hard drive, at least in the straightforward way. And I was also
00:08:06closely watching it whilst it described what it was about to do. So I let it do its thing
00:08:11and it ran a bunch of stuff here. And in the end it was done. And indeed it did successfully
00:08:18compress this document and made it significantly smaller. Now that's just a simple example maybe.
00:08:25And there would have been alternatives. My point just is it did that all in the command
00:08:29line in the terminal in the end by using our commands, our programs there. And of course,
00:08:36that all makes a lot of sense because we're talking about programs using a computer. And
00:08:41all these graphical user interfaces and rich websites were built for humans, for you and
00:08:46me. And that won't go away of course. But if we want to have little utility tools, AI agents
00:08:53running on our system that can at least do some of the tasks we are doing right now, then
00:08:59we need to give them a way of using the computer in a more efficient way. Because a graphical
00:09:03user interface, an app or a website built for a human is not the ideal way for a computer
00:09:09program of course. It would have to take a screenshot, figure out where the buttons are,
00:09:13move the mouse to a button, click that button, take a screenshot again to see what's on the
00:09:18new page. That's super inefficient, burns a lot of tokens and takes super long. And I mean,
00:09:24that's why we have the concept of APIs way before the advent of AI agents and large language
00:09:31models. Because if we are writing a program, doesn't matter if it's a website or an app.
00:09:37If we're writing a program and we want to interact with another program, with another service,
00:09:43of course in the past we already used an API and we didn't try to write a script that uses
00:09:49a website that's meant to be used by humans. That's why APIs exist and CLIs, command line
00:09:56programs, in the end are just wrappers around APIs, at least in the case of CLIs like the
00:10:03Google Workspace CLI. But that is exactly the kind of program we need and want for an agent
00:10:10to consume because it doesn't care about pretty buttons or anything about that. It wants a
00:10:15simple way of invoking various commands to get stuff done. And that is why this makes
00:10:22sense. That's also of course why we have markdown being more important now than ever and why
00:10:28most documentation pages already offer a little copy button like this, which makes it easy
00:10:32to copy the content as markdown so that you can paste it into your favorite large language
00:10:38or chat session or coding tool. Why some websites also support stuff like adding .md at the end
00:10:46of the URL to get this article in markdown because we're going towards a future where
00:10:52at least some services and some content will primarily be meant to be consumed by agents.
00:10:58I mean, take the documentation of a library or a framework like TanStack Start. If you're
00:11:03building a TanStack Start site these days, and of course doesn't matter which tech stack
00:11:09you use, you get my point, then you will likely do that with help of some coding agent, cursor,
00:11:15whatever. And if you want to tell those agents how to use the library, if you want to point
00:11:20them at a specific documentation article, you don't want to point them at a website like
00:11:25this. You don't want them to download the HTML code, which burns a lot of tokens unnecessarily.
00:11:32And that is kind of the same reason or the same reason for why CLI tools are becoming
00:11:38more and more important because we're moving towards a future where at least some tasks
00:11:42will be done with the help of AI agents or exclusively by AI agents. Which of course also
00:11:49means that if you are building some kind of service which is not primarily meant to be
00:11:54consumed by humans, you wanna strongly think about building a CLI as well as offering an
00:12:02API and whatever you need so that in the future, people can consume your service through agents.
00:12:09And of course, we're still super early here. The vast majority of people doesn't care about
00:12:14agents at all. And it's too early to tell how good AI agents will become and which kind of
00:12:20tasks they will be able to tackle in the future. Maybe we are kind of stuck at the current level
00:12:26where they can do some stuff, but definitely not all of that and still need human supervision.
00:12:31But even in that place, there are tasks that can be performed by agents and you can make
00:12:37them more useful and more powerful by giving them just the right tools that make it easy
00:12:42for agents to interact with our services, with websites and so on. And that's why we're kind
00:12:49of going full circle. Obviously, that does not mean that the graphical user interface
00:12:55and websites are going away and there will probably always be apps or websites that are
00:13:01meant to be consumed by humans that don't really make sense to be consumed by agents. I mean,
00:13:07something like Netflix. I don't see a huge advantage in an agent telling me what a certain
00:13:13movie is about. I guess I want to watch it. But for many services, especially in the SaaS
00:13:21business or in the professional services area, that definitely is the way forward. I think
00:13:28obviously early days, but definitely a clear development we can see here. At least that
00:13:34is my opinion. But as always, I want to find out what your opinion is, too. So please share
00:13:39it. Let me know what you think of that, what I maybe forgot or overlooked. And yeah, let's
00:13:44see how the world of CLI tools looks like in a year or two.