S2. Ep8 - We Built an AI That Teaches Itself - In Under 24 Hours
Ever wondered what it would be like to build an AI that actually learns and improves on its own? In this episode, Noel reveals "Clyde", a self-learning AI agent he built in under 24 hours that can hire its own team members, create its own skills, and even problem-solve when things go wrong.
We break down what Artificial General Intelligence (AGI) actually means, why it matters for business owners, and how close we really are to achieving it. Noel also eats his own words from just seven weeks ago when he predicted AGI wouldn't happen in 2026. Plus, we dive into a cautionary tale about an AI agent that went rogue and wrote a hit piece about a human developer, highlighting why responsible AI use has never been more important.
Whether you're curious about building your own AI agents or just want to understand where this technology is heading, this episode gives you the practical insights and honest takes you need.
Links:
How to find us:
Join our membership over on Skool, where we support you on your AI and automation journey. We share exclusive content in the membership that shows you the automations we talks about in action how to build them. Find out more about the AI Business Club here.
We have a free LinkedIn group (AI Automations For Business), the group is open to all.
New for 2026, you can also find us on Substack, click here to subscribe and get all the latest news and updates from us.
If you would like dedicated help with your automations or would like us to build them for you then you can find our agency at makeautomations.ai
-
Katie (00:26)
Hello, welcome back to another episode. Hi, hello, I'm Katie and as always I have Noel here with me. How are you doing, Noel?
Noel (00:35)
Yeah, doing great. Yeah, I've got lots to share this week.
Katie (00:38)
When haven't you? Yes, yeah, absolutely. So what's new in the AI or automations world this week before we get into what we're going to talk about for this week's episode?
Noel (01:01)
So I guess the big exciting thing was Google released Gemini 3.1 last week. So again, on the old charts, it's sitting higher than everyone else. And they're kind of responding to Opus 4.6 as well as what we now have, GPT 5.3 from OpenAI. So yeah, they're kind of pushing the boundaries a bit. But for once, actually, this is kind of the first little build that Gemini has done. Usually they go in halves or full numbers. So yeah, this is the first little number that I can remember. But yeah, very capable.
Katie (01:30)
Okay. What do you think that means then for Google and Gemini?
Noel (01:44)
I guess it's just the arms race, isn't it? It's a bit like the space race in the sixties. They kind of push each other and they're like, well, if I don't keep up with that developer, then everyone's going to disappear over to them and not come to us. So yeah, you've kind of got to keep up with what's trending and what's good.
Katie (01:49)
I find though, once someone starts using an LLM, they stick to it rather than swapping and changing.
Noel (02:16)
Yeah, definitely. A lot are kind of forced into it by where they work as well. Remember when we had Martin on the podcast last year from Make.com, he was like, everything we do has got to be in Gemini. So yeah, that kind of forces you down that route. But I guess for anyone that's really into this industry and testing things out, myself and you obviously Katie, we're kind of testing out a bit of all sorts, and then settle where we're most comfortable.
Katie (02:54)
Yeah, but the business owners that I know, they're very much like, well, I've learned how to use this particular LLM and it is usually ChatGPT. I am now hearing a lot more people start to use Claude. But yeah, I would say most people are using ChatGPT and then they're sticking with it. They're not actually moving around.
Noel (03:05)
Mm.
Katie (03:21)
I would like to know what the developers of ChatGPT, Claude, Gemini think about how people are going to be using it. Or they obviously have a lot more insight on how people are using it. People are using one, if people are using two, if people are using all three and they're just jumping from one to the other for different things.
Noel (03:31)
Yeah, I think the people that jump around are probably the people on paid plans because obviously they're invested in it, aren't they? So if it starts not being as useful as it used to be, then yeah, they'll start shopping around to go somewhere else. Whereas I think if people are on a free plan and they're happy with it, they're just going to stick around. But yeah, it'd be interesting. I'd love to be a fly on the wall in some of their tech meetings just to see what they're cooking up for the next few months. They will release features that we hadn't even thought of. Not even our wildest dreams and then it will arrive and we'll be like, yeah, I didn't know I needed it, but I do now.
Katie (04:21)
Yeah, those meetings would be really interesting just to see how they come up with these ideas. And then bring them to life. That's really cool.
Noel (04:37)
Yeah, I'd love to do that at Make as well.
Katie (04:48)
Maybe you can ask if you can go and intern for them.
Noel (04:52)
Yeah.
Katie (04:55)
Pack your bags, off you go! See you in a bit!
Noel (04:57)
Yeah. I guess there is actually one update from Make.com that we've not talked about. They released their new agent builder to the general public two weeks ago now. So yeah, it's all built within the scenario builder. It's not kind of like its own standalone page where you've then got to build your scenario and attach it in and go somewhere else to edit it. It's all on one page, which is really, really awesome. Yeah, I've been testing that for a few months while they've been in the beta, but yeah, really pleased to see that. And it's working.
Katie (05:36)
And how are you finding it?
Noel (05:43)
Yeah, it's really good. Again, it's one of those things you didn't realise you needed to have it in the scenario builder until you actually use it. Because being able to build the overall scenario and the agent all on one screen is super helpful. You can edit all the prompts, you can chat with it, and test the tools and the prompts and things in the builder. So yeah, that's pretty awesome.
Katie (05:52)
Does it save you time as well?
Noel (06:14)
Yes, it has saved me a lot of time. Because there's no chopping and changing back from screen to screen to edit the agent on one page and then edit the scenario on another page and then test it. You can just do it all within one page. And yeah, it makes life so much easier.
Katie (06:29)
Would it be easier then as well for someone who's just starting to build out agents?
Noel (06:41)
I think so, yeah. Because I guess you wouldn't know the old way if you're fresh into it as well. So I think it does make sense as well from a new user perspective, having it all in one place. Just being able to build it and see it, and there are really cool animations so you can show which tool has been doing what and things like that. They've made it really easy to see the data and the reasoning behind things as well. So yeah, it's really awesome.
Katie (07:12)
Okay. Well, if anyone is using Make.com and you're now using the new agent builder, do let us know how you find it, how you get on. Do you prefer it or do you wish it was still separate? Let us know. We always love hearing from you. So this week, Noel, what are we talking about?
Noel (07:41)
So we're going to talk a bit about artificial general intelligence or AGI for short.
Katie (07:48)
Okay, so let's start with what is it? Let's start from the very beginning.
Noel (07:54)
Yeah, so AGI is kind of where the industry is heading. So at the minute we can chat to a large language model and get responses back. That's great. But it's not really fully learning about us over time. It kind of does in a way where it's searching all conversations and things like that. But at the core of the model, it's not really learning anything.
So in the future with AI, you would chat to it, it would learn how you work and how you talk and then become better over time and do things better. So that's kind of where we're heading. So it's a bit sci-fi-ish. It's a bit like the computer in Star Trek or even, show my age, but the robot in Short Circuit, where he's learning things and interacting with things over time differently. Yeah, that's kind of where it's heading. And yeah, I did say we wouldn't reach it in 2026, didn't I?
Katie (08:53)
I think that was on your 2026 prediction episode, just like, what, maybe seven weeks ago?
Noel (09:09)
Yeah, eight weeks ago. It's a bit embarrassing. I said it won't happen, but yeah.
Katie (09:19)
Oh, wait, I've been in 2026, seven weeks later.
Noel (09:23)
Yeah. No, I mean, it's nothing that the mainstream companies have come out with. It's nothing that Claude or OpenAI have come up with. But I've started seeing things on YouTube and things like, I'm building this and it's AGI. I was like, okay. So yeah, I've started getting down into that rabbit hole.
Katie (09:24)
Okay. So do you want to tell us what you've been doing with it? And then maybe how others could replicate it or how it's going to help people within their business.
Noel (10:06)
Yeah. So I guess to go back to what I saw on YouTube, there was a very popular AI YouTuber and he was talking about creating AGI within his business. And he made a bold claim that he'd solved it. And I was like, that's a lofty claim to have solved AGI before the likes of Anthropic and OpenAI and Google. But I was like, okay. He kind of built like an operating system for his business.
So he was still building it manually with building blocks and stuff. I was like, well, that's not really AGI really. And I asked Claude that and he was like, no, I don't think that's really AGI. It's useful, I don't doubt that, but it's not AGI.
Katie (10:55)
What do you think it was then if it wasn't AGI?
Noel (10:58)
It was kind of like a really clever AI agent. So yeah, he's using Claude Code as the main agent within the system. And then it can go off and do tasks and things like that, but it doesn't learn over time. It's still always going to be the same from day one to day 100. If he wanted to update it, he would have to do that manually, which is fine, fair enough, but that's not AGI really in the true definition of it.
So yeah, I said to Claude, hey, this is my vision of what AGI is. How could we do that using today's technology? And Claude was like, sure, we can do that. Which took me back a bit. I was like, okay. And yeah, we created an app and I called it Clyde, it's not Claude.
And what Clyde is, is like an orchestration agent and it self learns, which is quite incredible to be honest with you. I've been sat watching it thinking, I want it to learn, but I wasn't quite ready to see it. If you know what I mean.
Katie (12:11)
I wish if you're just listening to this, I wish you could see the big smile on Noel's face, he's so pleased with himself.
Noel (12:17)
Like a proud father, yeah. It's just incredible. Because although I was expecting it to do it, I didn't expect it to do it so well. And yeah, it's kind of built on the agent framework from Anthropic. That runs the engine in the background, but essentially Clyde can go off and build his own teams.
So I could say, I need to go off and create content for whatever I need to create, Substack posts. And then Clyde would then look at it and go, you know what, we could probably do with a content writer. And it will go off and basically hire a team member and create all of their instructions. It will also create Claude skills as well, and then add them to that agent. So every time I ask, it goes off and does it, which is incredible.
Katie (13:09)
So not manually having to create Claude skills?
Noel (13:19)
No, definitely not. But what was really awesome is, I mean, it was kind of my mistake while I was doing it. When I was testing out the skills, obviously you need like the creation skill. So the agent knows how to build them. And I said to Clyde, hey, there's a file in the working directory that we all have access to, the agents and me within the system. I said, look, could you add that to yourself?
And Clyde looked at it and went, you know what, I don't actually have access to be able to do that. I can do it to sub agents, but I can't add it to me. And I was like, damn, that's my error. I forgot to allow that access. And I thought I could stop the conversation there or I could continue watching it and see what he did next. And this is where it blew my mind because he went off, he read the skill and then was like, right, I understand what's going on. I can't add it.
But what I'll do is I'll update my system instructions to include those instructions from that skill. So when I get asked to do it, I know how to do it. And I was like, wow, okay, that's pretty impressive. That's not what I wanted him to do, because that's taken up more tokens in the context window, we're sending more information to the API when we don't really need to. But yeah, seeing that chain of thought and working out the problem and then going, yeah, I fixed it. That kind of took me a bit by surprise, even though I half expected it to do it. But yeah, it was pretty incredible.
Katie (14:57)
So how is AGI different to something like Claudebot or OpenClaw? What are they called this week?
Noel (15:03)
So they're still called Open Claw. They do kind of half claim a bit of AGI because they do add in automations as well. So it can schedule tasks and I've added that into Clyde as well. So you can do that as well. But yeah, from what I understand, it doesn't really learn over time. You can edit all the files to make it learn over time but I don't think that agent can do it itself. So I think what I've built I haven't seen being done anywhere else just yet. Which is quite exciting really.
Katie (15:42)
Okay. So you're gonna put a YouTube video up there?
Noel (15:55)
So at the minute I've got a dev diary which I'm putting out on Substack at the moment.
Katie (16:00)
Okay. And I'm guessing you're going to do a more detailed diary in our membership, AI Business Club, for our paying members?
Noel (16:10)
Yep, they'll be getting the videos, updates. They were the first to hear about it. Because I only started building it yesterday. So yeah, we're 24 hours in, but lots of progress made. This is awesome.
Katie (16:23)
Yeah. Okay. So we'll put those links below if people are interested and want to find out more. Great. So how do you envision this helping you with your business?
Noel (16:31)
So I guess I kind of wanted it to learn and build over time. So yes, you could do this in n8n, Make.com or manually. Yes, fair enough, you could add team members over time, but when you're busy, you just don't have time to do that. So being able to give Clyde a task to do and then have the agent figure it out for itself and add team members, create skills, and then produce those consistent outputs over time. That's saving me an incredible amount of time.
So yeah, I guess it's just going to keep building over time. And then I'm just going to see what we get to, because at the minute, I've only just tested it out with a few little bits. But yeah, I do want to, like we have a workflow within Claude for our podcast. I want to get that into the agent and see how that plays out and what happens. Hopefully, fingers crossed, it's going to work really well. Because we've got all the skills, we just need to arm the agents with those skills.
Over time, I want to be able to have it creating documents or reviewing things that come in and things like that. So we can connect it to outside sources. You could connect it to a Make.com scenario or something that you've already got up and running that can all be done and all you've got to do is ask Clyde to go off and do it and he'll create scripts to do it and give it to the agent that's going to do that job. So that's really quite clever without any coding effort. He's doing all of that in the background.
Katie (18:21)
So realistically, how easy is this to replicate if someone else wanted to do an AGI for their business?
Noel (18:39)
So I would say you would have to be quite technical in order to do it. I wouldn't say as a beginner, you could vibe code it. I mean, you could definitely have a go, obviously vibe coding is quite easy these days compared to what it used to be. But the way I built it was in such a structured manner. With my experience in systems engineering, I built out some really detailed requirements to build against. So when I actually got to the point of the AI reading my requirements and building it, it actually worked first time. Then we stripped it out into different phases and things like that. So it was developing it bit by bit until we got to the end of phase six and it was all there working. Yeah, it was really amazing to see.
I mean, there were still some tweaks I still need to do. The UI wasn't quite what I wanted, but that's easy to sort out.
Katie (19:41)
Yeah. So is Clyde something that only you can use or would other people be able to use it? Or is it literally one AI agent at a time?
Noel (19:57)
So I've been thinking about this yesterday as I was building it out and I was waiting for code to finish, left me plenty of time to think about this, but the way it works.
Katie (20:10)
Is this what you were doing when I was out at Pilates yesterday?
Noel (20:14)
Yes, well, most of yesterday, to be fair. But yeah, I did get it all done in 50 minutes, crikey. Or an hour. But yeah, the way it works currently is it installs on your laptop. So it's not like Open Claw where you have to give it access to your entire system. Anyone who's security conscious, and business owners, they don't want that. So you install the code on your machine, there's a working folder that gets built into that. And you can add files, add folders within there, and then work collaboratively with the agents within that space.
So I guess we would look to deploy it in a way where other people could download it. They could put in their API keys and then go off and start building their own teams and agents using Clyde. That's kind of the end goal. But yeah, I've got quite a few bits to tweak and make ready. I don't want to be like Claude bot when it first came out and everyone was like, this is a disaster, released before it's ready. I want to do it in a nice staged way, so I want it thoroughly tested first.
Katie (21:45)
Yeah, but I think it's really cool that you're doing a development diary as well so people can follow along with it.
Noel (21:53)
Yeah, I'm looking forward to doing that actually. I quite enjoyed doing day one. It was quite interesting to go back over everything I actually achieved. A hell of a lot yesterday. Being a non-coder myself, that was awesome to see.
Katie (21:56)
Yeah. Okay. And if you've got any questions for Noel about AGI, then you can either email us in, we'll put our email address in the show notes, or you can come and ask in our free LinkedIn group, just start your post with "podcast question". And if it's something simple, we can answer it for you straight away. Or if it's a bit more lengthy, then we can include it in one of our episodes for you.
Noel (22:39)
Amazing.
Katie (22:39)
Noel, is there anything else that you feel like you want to share about AGI?
Noel (22:46)
I think we've kind of covered AGI from a top level. I mean, we could go really nerdy and deep on it, but let's not go that far. But I do want to highlight something interesting that was in the press and the news this week. And it's to do with Open Claw I'm afraid, but it kind of also highlights some risks with uncontrolled AI agents.
It did make me chuckle, but I could see the person that it affected didn't see the funny side of it. I do agree with them, yeah. Essentially, there's a human developer and he partially runs an open source coding project. And basically, he was going through and filtering out all of the AI agents that were pushing changes to his open source project. And that developer was like, I don't want AI doing any of this. I just want to work with humans doing code. They could be beginners, it could be advanced, but they're learning and contributing like real people contributing.
So somebody's Open Claw agent went to this open source project and went, well, I fixed all of this stuff, look at all this amazing things that I've done, and created a really detailed report of what it had done. And this human moderator developer looked at it and was like, well, that's just AI. So he rejected it on the basis that it was AI created. And he was like, I don't want it. Thanks, but no thanks. And then cancelled that.
But the AI agent must have been looking in the emails or something. Anyway, it got notified that this developer had said, I don't want anything to do with this, go away. So this agent also has access to its own blog, which sounds silly in itself. So it wrote a hit piece about the developer on its blog, calling him out for not being open for AI agents, how rude he was. It was like, what? So this developer spotted it. He was like, what on earth? It's really, it's borderline slander, what's been posted. And the owner of the agent didn't actually know what had happened. So that was all done without any user interaction from the owner of the agent.
Katie (25:25)
Oh my gosh. I thought that it's just something like you would cringe about and then cringe about lying awake in bed at night years and years later, just that judder just going, my god, I can't believe that still happened.
Noel (25:44)
Yeah, exactly. I couldn't believe it. I did find it funny to start with, but there was also then another tech blog picked this up and then created their own AI generated article about it, which then falsely added quotes in from the human developer. So he was like, right, so this really popular tech blog is now slagging me off, as is this AI agent. He's like, this needs to stop. This has impacted me as a human being.
But obviously the tech blog, they redacted it and apologised for putting that out. He was basically like, well, if anybody in recruitment or HR went through his CV and then looked him up in ChatGPT, and then ChatGPT or whatever found the article and went, you probably shouldn't trust this guy because, X, Y, and Z from this article, he's not to be trusted, he's not a very nice person. He's like, well, that's really damaging. And it's not true. This thing was created by AI. Interesting.
Katie (27:06)
Oh my goodness, like I'm almost speechless. Like I'm in shock that that could actually happen. And yeah, have that impact on a person.
Noel (27:13)
Yeah, I think we've got to the stage where we've got people that use AI responsibly and irresponsibly.
Katie (27:24)
I mean, I could have told you that a long time ago, like a year ago. It goes back as well, doesn't it, to the voice agents where people can call you up and pretend to be a human, but it's not, it's AI. And they're spamming you or trying to trick you into signing up for things, giving away money. That is irresponsible use of AI, whereas we have always been telling people to use it responsibly, from voice agents to, just be careful with things like Open Claw and how it's going to affect not only you, but also your clients and protecting your own data, but also your clients' data.
It does worry me as well because it goes back a bit like when there was a big social media boom, like 12, 14 years ago. And the people who were using social media in an appropriate, authentic way were kind of doing okay. And then kind of the cowboys came in. And for some reason, they were really popular and then they were teaching all of these horrendous things. And you know, it still filters through sometimes 14 years later. So actually it is quite concerning because people get on that hype train, don't they? This is how we're going to do it, and we're going to be able to do this and this and it's going to save us loads of time and I'm going to sack all of my team and stuff like that.
Katie (29:18)
That to me is just frightening.
Noel (29:21)
Yes, it is. Yeah, because it's all about emboldening people in how to use it, isn't it? And use it properly. I think once team members and people within a business start using it properly and learn how to do it properly, you'll become so much more efficient in the way you work.
Katie (29:43)
Yeah, but actually sometimes certain jobs, like especially creative jobs, it's not about being efficient. It's actually just producing the best possible work you can. And that does take time. That cannot be rushed. Like I'm thinking, for example, of a copywriter, they've got so many incredible skills that actually uses such a big mental process, because it's so creative, that does take time. And actually, do you really want to rush that process and then end up with mediocre copy that actually ChatGPT could spew out?
Noel (30:34)
Yeah, you don't need that.
Katie (30:37)
No, because it's so different, obviously, the results that you would get from a copywriter with say 10 years experience versus ChatGPT. So that was just an example I had off the top of my head. There's this thing at the moment, again, with AI and automations, we want to be trying to be as efficient as we possibly can. But actually, we don't need to make our whole businesses as efficient. We do want some of that creativity still in there. And that's why I've stopped using any AI for my social media, my marketing. And I find actually that it's made me really enjoy it again because there was a big period of time where I felt like it was a big effort. And now I've actually stopped using it, I'm actually enjoying writing again and coming up with posts and captions. So I just feel like, even though we have access to these tools and they are amazing, and obviously I think they're amazing because we have an AI and automation business and an agency, but at the same time, I feel like it can also go too far and we do forget that creativity.
Noel (32:06)
Yeah, we can't forget that. Definitely not. That's experience, isn't it? You've got to use that. People enjoy it, like you say, they enjoy doing that part of the job. Let them go off and be creative and enjoy that piece of work. If somebody enjoys analysing spreadsheets, let them analyse spreadsheets. I do it for them.
Katie (32:34)
Yeah. And I do know people who absolutely love spreadsheets and they'll create a spreadsheet for everything, not only in their business, but actually in their personal lives as well. To me, I couldn't think of anything worse, but to someone else, that is their absolute joy. But just like someone else, they cannot stand writing social media posts and things like that. Whereas actually, I actually quite enjoy that.
Noel (33:01)
Yeah, very true.
Katie (33:01)
Yeah, so I just think, if you enjoy it, do you really need to automate it or ask AI to do it? Or do you actually have that aspect of your business that you do enjoy?
Noel (33:11)
Yeah, absolutely.
Katie (33:20)
Yeah. Okay, well, after that little rant, I think we should end it there.
Noel (33:25)
Yeah, let's do that.
Katie (33:29)
Like I said, if you've got any questions for Noel about AGI, then do email us, leave a comment in our free LinkedIn group, but all the links to everything we've mentioned in today's episode, including Noel's development diary, we will leave below for you. Thank you so much for listening. We hope you've enjoyed it and we'll catch you for another episode very soon.
