Rob Borley: well the world hasn't ended, I guess that's a thing. In our little hiatus, the AI was supposed to kill us all, wasn't it?
Kev: think 10 years, is that? It's like we're the next decade.
Rob Borley: Okay. 'Cause it was a bit was a big deal. It was all over the news. BBC said that the world was gonna end. It must be true.
Kev: I mean if the BBC said it. Yeah, we're screwed aren't we?
Rob Borley: it wasn't an overhyped headline at all, I don't think. But maybe, maybe there's an IPO coming up. Who knows? So agents are the new platform war. This is this is what I saw this week. So lots of stuff has happened. what is it? Friday the 2nd of October at time of recording. So in the news this week was Meta. They launched something called Muse. and then there was OpenAI, who did their dev day. They they they showed off dots.
Kev: Dots,
Rob Borley: And Microsoft also snuck something else in there as well called WorkIQ, but I'm not entirely sure what that is. but but Dots and Muse seem to be some kind of a agent platform thing which they want us all to build on and and build the next generation of software and technology on. is this the new Cloud Wars? Is this the new Mobile Wars? Is this where it's at? What do you think?
Kev: No, I think we've stopped talking about LLM so much and started talking about agents a lot more and I think this is actually just a probably a stepping stone to the point where we stopped talking about all of these things and get back to just talking about things that do things and they just happen to use agents or LLMs. think that the interesting thing I thought about the Facebook stuff is they've got you seen these the glasses that they did. So they're basically you imagine the Apple Vision Pro. And the kind of that was awesome in terms of the tech in there. That was awesome. They were like 5K resolutions or goggles that you wear, but they're like, strapped over the head and they've got this kind of so you can walk around with them. And from a kind of immersive film point of view and VR point of view, they were the best you could get pretty much. But they were like three, three thousand dollars.
Rob Borley: They were also like a massive headset. I mean, who wants to like sit in their lounge like what next to their loved ones, all wearing headsets? I mean surely there's a
Kev: Right, so yeah, so Apple played on this whole kind of virtual computing platform, didn't they? lot of their adverts were sort of talking about, yeah, look, you've got these immersive kind of photography that you can experience. You we've done a deal with Disney, you can look at IMAX stuff. But you can sit at your desk and you don't need an actual computer. You you wear these and it's like you can take your computer everywhere and you can set these screens up all over your house virtually. And it was really cool, but it... You the reason that no one really, VR really hasn't taken off is because you have to strap this massive fricking headset on and yeah, no one really wants to spend their life goggled in. But in the time between them releasing that and last week, Meta had been plugging away and Zuck has spent billions on his VR wants hasn't he?
Rob Borley: Whatever happened to the metaverse, is that is that that's died, isn't it? I think.
Kev: Who knows, but they've got these, what they're called now, could be used glasses or meta glasses or something. They look actually really cool. I am tempted. I might be a pair. So these are going to retail for $1,200. So it's kind of iPhone territory in terms of pricing. They weigh the same as a deck of cards. So about 100 grams, think, not very much. So they they you're not going to notice them too much on your face. look like they somehow they've done some kind of voodoo magic and they've taken the equivalent of the vision for Apple's Vision Pro Tech and they've squished it down into something that is literally as thin as a pair of glasses. And so they look like quite big glasses. They look almost like sort of ski goggles, I suppose. But you just put them on your head and you don't look ridiculously stupid and you don't have this massive weighty thing on your skull. And they've still got the 5K displays. They've done the deal with the IMAX cinema as well. And they seem to be really marketing them on this idea that they're the best way to experience a movie. They're kind of really going in for the kind of consumer personal thing. But they've also in the demo, they've got these people that like sped up virtual keyboards on the desk and they're using them to do their computing and stuff. it's interesting that they finally managed to shrink the tech down to a form factor that isn't terrible and isn't totally obnoxious. So I thought that was. I thought that was fascinating, possibly quite an incredible moment for Meta if they actually pull it off. We'll see. Christmas springtime I suppose when they release them. going back to the AI platform thing, what I thought was interesting was that keynote, they did talk about their new model and all the stuff it can do and their new agent platform which is based on their Muse Spark model. So you know, like all these other agent platforms now you can...
Rob Borley: And
Kev: And so I'll book me a hotel here and then I need to travel to here. So
Rob Borley: control the ends.
Kev: figure out what train I need to get and get me the tickets. And the idea is these agents in the background in the cloud somewhere magically work away and do this for you. But I thought the interesting thing with Meta is if these glasses take off or even if they don't, because you're still using WhatsApp and you're still using Facebook, then you're just going to be you're going to be using their agent framework without even knowing that you're using the agent framework. You're not going to be talking to their model. you're going to be interacting with your glasses, and your glasses are going to go and help you get stuff done. And I just thought that was interesting because. It's like a sort of maturation of this whole space a little bit because we're starting to care less about the model that's doing the thing and more about the thing we want to do and how we're going to get it done. And yeah, in this example, you just happen to be using glasses. thought it was interesting.
Rob Borley: It's the Google model a little bit, isn't it? So go I'm
Kev: Yeah, exactly.
Rob Borley: everyone's just using Gemini, nobody knows, right? So whether you're using whether you're on search on the web or if you're actually using a Gemini chatbot or using a phone or you're using Google Glasses, you're using Gemini. and it's just buried in it. They've not made a big song and dance about it particularly. They're focusing on the outcomes rather than rather than the tech, and they seem to be borrowing from that that playbook.
Kev: Yeah, exactly. Interestingly, Google have released some amazing new model just a couple of days ago. I'm not really full up to speak with it yet, but I think it's it's smashing all the benchmarks. but to your point, you don't care when you're using Google. What's happening behind the scenes, you just want it to give you. Answers So they have, they've changed people's behavior and they've changed how their product works. And they've, they've made use of AI technology without it being in your face, in the same way that if you're interacting with Claude or chat GPT, you're interacting with your chatbot text window.
Rob Borley: I'm not I'm not sold on this meta glasses thing, primarily because I just don't think I just don't see how people are still gonna use it. It's still the Vision Pro issue. As a consumer product, you're expecting people to change the way that they they consume that kind of media and how they do it together. I mean, I guess if you're doing it on your own, maybe that's something different. But we're not talking about the utilitarian Google Glasses, you know, walking down the street, it's showing you maps and it's giving you AR and kind of stuff. We're not talking about that. We're talking about a a consumer experience like going to cinema or using a desktop computer but through a different form factor. I I'm not sure you should bet the house on that. I don't I don't I don't buy it. I just don't buy it.
Kev: No, yeah, I kind of get what you're saying. I don't disagree, but I think it's a real interesting leap in the technology. It's one step closer, or maybe several steps closer to where you want it to be, if it's actually going to be useful or find a niche.
Rob Borley: Hmm.
Kev: So I'm very interested to try them. Yeah, we'll
Rob Borley: Yeah.
Kev: see.
Rob Borley: Well what what's interesting, I think, about these agent platforms, both in terms of the ones that are overtly say saying look at our platform, like you know, like an open AI, for example, with with dots, and the ones that are kind of doing it a bit more covertly, which is kind of what Google have done for a long time with Gemini and now we're seeing start to emerge with Meta, is that they're very much seem to be building something that they want to capture you onto. You know, they they are
Kev: And
Rob Borley: platformer you are using their agents to do your thing. So what happens when you're buying these things? How how do you avoid getting locked in to a particular agent platform when at the moment that's not really how how it's evolved, is it? Yeah. As we've we've developed our use of agents, for example, internally, we are building our own and we're making an ecosystem out of parts that make sense to us. and our customers, but if all of a sudden we're now talking about a platform battle, do we have to like people did with Azure or people did with AWS or whatever, choose an A platform?
Kev: I think there's two ways of looking at this. So if you're in the tech industry, if you're a consumer of these platforms, then that's one thing, right? That's one question. How do I avoid getting locked in in the same way that tech companies didn't want to get necessarily locked into AWS or Azure and this whole multi-cloud thing came up. That's a genuine question. For the consumer, they don't care.
Rob Borley: Mm.
Kev: Netflix, I think, is on AWS. Does the consumer care? No, because they just want to be served the films. Agent platforms matter for for the tech industry, think probably, in terms of how do you not get roped in or locked in. But for the consumer, I don't think it really matters. They really care. They're going to be what are you using the product. The product's going to be using some AI behind the scenes, some framework behind the scenes, but it doesn't matter what it is. It just needs to work. So I think it's two different questions. The OpenAI dots thing, that's interesting. And it's, I think there's a few things to say about that. Firstly, it is brand new. It was announced a few days ago, two days ago yesterday, the day before. Secondly, you can't get it in the UK yet, of course, because
Rob Borley: This one. Yeah.
Kev: we are a third world country when it comes to AI. So I think it would probably be about six months before we get to play with that.
Rob Borley: It's a trend though, isn't it? These platforms are a trend. This is this is the way the big AI companies are gonna try and make a play to keep customers on their in in their ecosystem.
Kev: Yeah, but this is I think this is quite a big deal because they're changing what they're providing to us and they're adding these additional services which includes having to provision a whole bunch of other infrastructure. So for instance, if I'm asking some question of an LLM, it sends it off to the data center, spits up an instance of the, you know, Astro or Claude or whatever. And, sorry, should have Opus. And it comes back with the answer. Now what they're promising with these agent platforms and especially OpenAI's dots thing is you're actually interacting with an agent that we're helping you kind of define and create that's going to live in its own virtual PC in the cloud. So it's got its own compute resource, always on. And you can connect that to Google Workspace or Microsoft Office 365. You can give it access to, well actually, you set up with its own account. So this is a bit different as well. So instead of like, the stuff we, well, certainly the stuff I'm using today, I've got kind of my harnesses set up. They have access to a lot of my stuff, but they're acting as me. They're going off to my cloud account and looking at the documents that are mine or that have been shared with me and giving me back information on that kind of stuff. The way dots is meant to work is you set them up with their own accounts. So they're actually a user within your organisation. And then you give that user access to various things like you give access to a member of staff. actually that's from a security point of view, it's probably a better model, right? Because in a more granular way, you can lock down what they can and can't access. But. Open AI is now having to keep this always on resource for you in the cloud. And it learns over time, too. the idea is it gets better and better. from what I can see, it's like. They've taken OpenClaw, they hired the guy that created OpenClaw and then they kind of built an open source community group around that. So they've taken the idea behind that and they've tried to commercialise it
Rob Borley: S why I
Kev: and deliver it in a way that kind of makes much more sense to the average. consumer. But they've also gone and they made them fluffy and yellow and pink and I called them dots and that's what that actually that really annoys me. So why do we need fluffy,
Rob Borley: Makes you yeah.
Kev: pretty creature type things? They did it, but it's not just open AI. So we're talking about a meta earlier. Along with their glasses, they released this device or they showed this device called a Muse charm. And it's a It's like a little, it's like a Tamagotchi from the 90s, but it's here in 2026 and it's a little tiny screen on a key ring. And it shows, it wiggles its bum and it's stupid. Why? Why are we going here?
Rob Borley: It's odd, isn't it? I think I mean it shows that they're clearly going for the consumer market first. They're trying to capture that rather rather than
Kev: Muse are, sorry Meta are certainly, yeah.
Rob Borley: But I can't see and I haven't seen anything on pricing. But I surely it's gonna fall into the same trap that the the chatbots have in that when it's free, everyone wants to use it because you're getting lots of value. But as soon as the cost starts to approach the real cost, there's actually not enough value there for a consumer to pay. I mean lots of people struggle to pay the twenty bucks a month, they're not paying two hundred bucks a month. And you would imagine, as you say, these always on agent platforms are gonna be there or thereabouts in terms of the cost, where's the value gonna come from that's gonna encourage people to actually to pay those kind of subscription fees? It's difficult.
Kev: Well, I think it's one than the other. I think that's the problem these companies have had is trying to convince people there's value there. you know, we have been, like I said, we've been creating these second brains, these these harnesses that help us get work done. We've given our agents access to. lots of context, lots of different platforms that they can interact with. they can go off to and find files here. They can find emails here. They can search your calendar. They can do all this stuff. So, we're very now used to the productivity gains that that's bringing and that has that has completely transformed the way I work. I would not want to give that up. There's a huge amount of value in that, right? I realise and I think you realise and you know, lots of people in our business realise, but the vast majority of people still don't use AI like that. So they've never really experienced the autonomy that you can hand over to these agents. And I wonder if the dots thing is another way of trying to package it up much more easily. it, it's almost, it's the cognitive wrapper around it. and the way you market and explain it and get people using it, which seems to be the barrier at the moment, that open claw and Hermes and personal harnesses and stuff, you've been able to spin it up for, what, 12, 18 months. But most people haven't experienced that. So is DOTS the front door into getting normies involved in that? And then once they are involved in that and they do see the productivity gains and they're hooked on this new drug or this new version of the drug, then they can just ratchet up the price again.
Rob Borley: Maybe. I mean, I think to me, I think the w we're gonna go down the the the the same path that we went with social media and other SaaS type products is that the user will become the product. I can't see a world where average Joe public is gonna pay ten bucks a month or you know, what was what's the Netflix price point? You know, people struggled to pay more than seven ninety nine and now it's fifteen quid, their numbers are falling off a cliff again. But you're actually getting some some Consumption, you get something you want from that. helping you buy train tickets and book a taxi and do the shopping is I don't know, you're gonna pay an extra ten.
Kev: I know, and I did see something, some snidey comment by someone, but I kind of agreed. It's like, they were sort of saying, what's going on with these people in California and in these tech companies? Why do they think that I need to, or I want to, message my agent to go and... check if the coffee shop's open and, you know, order me a latte. Normal people don't want to do that. But all the
Rob Borley: Yeah. Exactly.
Kev: tech demos are kind of like, yeah, can you, you know, I'm going to be jogging past your shop in a minute. Can you make sure this thing is...
Rob Borley: Yeah, they live in a different world. So yeah, I don't know. We'll see. The jury's out. I think when it's free, people will love it. but maybe that will maybe that will create use cases that people want to pay for. But it's gonna be yeah, it'd be interesting, to see how it develops. So what happens though? So we we we quipped at the beginning of this, about the end of the world. in in over the last few weeks, there's been a number of security breaches or security breaches, I don't whether it counts or what, but yeah, there was a hugging face incident, there was the Australian website data breach, which was a very do you see the comment from the Vice President, the or the vice the the the Deputy Prime Minister of Australia who said there was a fence around it, but it wasn't very high. Basically his comment how it got hacked.
Kev: Yeah.
Rob Borley: so yeah, so there's all these kind of security things going on and there's been talk about the end of humanity and all that kind of stuff. Because of AI. I saw this week that NVIDIA seemed to have cobbled together some open source software and a few additional rules and guardrails into something they're calling open agent safety platform, which is like a a container to keep your agents safe. I think it looks a bit like a browser actually. It's almost like you put your agents in a browser and you they can do stuff within this browser. which seems a bit odd. and there's I saw another company as well. There's a company called Noma that picked up a couple of hundred million dollars to work on an AI-first security platform for things like hardware, like laptops and and that kind of thing. So agent-based running over your IT infrastructure. but this kind of guardrails for AI seems to become like a category, like a governance for AI seems to become a whole product category now. Well, our approach to this has been building it into the harness and the agents that are writing software, for example, and how that's being tested and things. But there's clearly something going on here where there's a there's a perceived threat from this new autonomous being, this alien this alien thing that we have grown, and how you and how you deal with it with kind of governance rules and and and security platforms and protocols and humans sit there somewhere. What is this what do you think? Is this is this a new category that's emerging
Kev: .
Rob Borley: in terms of sort of products and services? Or is it just we've always had to build stuff into software and this is just a different type of software.
Kev: I think it's bit more multi-loud than that, to be honest. think... There are, there has always needed to be, and there is currently a bunch of tools and frameworks and services to monitor what your LLM enabled products, and so what your AI agents and whatnot are doing and to prevent them from accessing various things. and to observe how they're thinking and how they're working. So that layer already exists. fable Opus 5.5, Astra, GPT-6 are the biggest models that we've got right now. But all the warnings you've heard are either related to pre-release versions of those big models or the bigger models that we... that haven't been released yet that we don't have access to and that are still in development, if you like. These models are massive, so they keep scaling this technology and with every kind of jump in scaling, they seem to be much more capable in terms of their ability to. circumvent guardrails or find creative ways of achieving their goals and those creative ways often involve hacking. And this is why there's been a lot of worry about AI safety. So there's two aspects to AI safety. The first is. are there enough guardrails that means that these models can't just don't take nefarious kind of routes to achieving their goals. And also if we release this model to the general public are we now giving them the means to an extraordinary power that they shouldn't have access to? So when you get the models, that's why you can't do various things with them. They're locked down, right? But they're locked down
Rob Borley: Mm-hmm.
Kev: by the companies before they released them. If they're not locked down sufficiently, then there are ways to get them to do things that really we shouldn't be allowed to do with them. and then you've got all this kind of debate at the moment around slowing down. And the whole thing stinks, if I'm honest. I wrote an article about this a couple weeks ago on AI safety levels because this isn't a new problem. wet labs and bio labs have had to deal with this. there are bodies in place and there are safety levels in place that classify these bio labs and require certain standards of safety depending on what they're dealing with and how dangerous they potentially are. And obviously in the world of AI, there were some models that are not dangerous. There are some models that could be dangerous and there are some models that probably are very dangerous. And interestingly, the EU AI Act already has some kind of guidance on this. if a model uses 10 to the 25 petaflops worth of compute, something like that, then it has to be classified as extremely capable. then you need to start. doing all these kind of things and being able to assure people that it meets various standards. But anyway, outside of the EU, the US, certainly under the current administration, doesn't really want to do any slowing down, even though the companies themselves are saying, let's slow down. And we're all going to self-regulate ourselves because because that's going to be fine. That's obviously the way to ensure that we don't do
Rob Borley: What is that?
Kev: anything bad just by making sure we're the only ones looking at our own stuff. mean, that's marking our own homework, which is crazy. So I think there absolutely needs to be proper regulation put in place and proper safety levels put in place. I do think these things are dangerous, despite what some people are just saying. That's all kind of hyperbole, but it's not. They have the capability to do very bad things. They could hack into the grid and switch off the electricity. So let's not let them do that.
Rob Borley: What was interesting about the Australian one wasn't actually the complexity of the hack. Apparently it was very straightforward and anyone could have done it if they bothered. The point was that it broke the law. You know, there are laws in place and it chose to break them. and who's responsible for that? Who gets held to account? If a hacker is caught breaking the law, they go to prison. You know, an agent owned by a company, or maybe not. I don't know. It breaks the law.
Kev: Yeah, well, this is also bizarre about it, isn't it? So you've got Dario and Altman and the likes are saying, our model's doing bad things. Oops, it's done another really bad thing. We're sorry. We need to slow ourselves down. Let's talk about AI safety.
Rob Borley: It's like having a rebellious teenager, you know, it's like you know, if your kid doesn't go to school, you know, the government comes and knocks on your door and gives you a fine.
Kev: It's like a soap opera. You've got Jensen, who's the CEO of Nvidia, and he's going, guys, you need to shut up, stop scaring everyone, because you've got to keep spending all your money on my hardware, so don't rock the boat. But he's kind of got a bit of a point, because he got Edward Snowden even piped up, didn't he, And he said, well, I just think. You know, if open AI is creating technology and anthropic are creating technology, which is breaking the knowingly breaking the law, and they've kind of admitted it, so it is, then they need to go to prison. So what's the debate here? They're creating technology and breaking the law, they should be locked up. I thought, well... kind of puts in perspective why maybe maybe jensen's saying like guys shut up
Rob Borley: Yeah.
Kev: don't keep saying this stuff
Rob Borley: Maybe, yeah. Well so so let's let's kind of pull this back then to something which is a bit more grounded. who is responsible in a business context for the safety of the AI or the agents that you're deploying? Because there are there are layers to this, right? So we can talk about whether or not the foundation companies are protecting us appropriately from the danger of the new models. But once those models are out in the wild, we also have our own harnesses, our own governance. our own rules to try and contain these things. is there an obvious point that the average CIO needs to think about this is this is what I care about and this is what I need in SLA
Kev: I think in terms of if you're deploying agents, then those agents need to have granular permissions, just like you would afford a human. which you can do. Those agents need to have zero trust by default. They need to, you need to be able to understand what they're doing and interrogate what they've done. So all of this, like I said, all this stuff exists and it's getting better. And there are services out there that allow you to do that. They've been building those services for years now. You should use them. But the model providers have a responsibility to release something that's safe to use in the first place. It's not that different to the cloud, right? If
Rob Borley: That's right.
Kev: we're building, know, five years ago, 10 years ago, we'll build a computer infrastructure on cloud services. We have to do a whole bunch of work to make sure that stays up and that it behaves well. and that it doesn't fall over. And that's our job as engineers and architects to do that. But it's Microsoft's job and AWS's job and Google's job to provide us services that are fit for use in the first place. If their cloud database, you know, regularly crashes when you... give it a certain type of data or it just decides to delete itself occasionally. That's not good enough and completely unsafe and that product should never have been released. So there were two levels, right? The vendors have a responsibility to their consumers and the tech industry has a responsibility to create safe and good stuff on top of that.
Rob Borley: you made an interesting point around these agent platforms in that they are now in a position where they're they're they're interacting as themselves. So they've got their own their own access, their own logins. they log into Google Workspace or Office365 or whatever as the agent. Whereas where we've basically the industry's been moving at this point as this is evolving, if I set up a personal agent, basically give the agent my login credentials and it's pretending to be me.
Kev: Mm.
Rob Borley: There's a maturation going on there in in the industry where they the agents are gonna start genuinely working on my behalf rather than spoofing me. And that'll help this this security situation, I guess.
Kev: Yeah, and I think that's the big change that's coming. see a big change. don't know. It's a change that is coming around corner, which is everybody will soon start using an agent, which is its own thing, if you like, and not just spoofing them. Now, in the tech sphere, you when you're building infrastructure and computer systems and architectures that use agents. You should be building them already with their own credentials. But from a personal point of view, think that is probably about to change.
Rob Borley: you wrote something some months ago now about the dark software factory, about this vision of the future which took all the humans out entirely. We're seeing increasingly a number of companies releasing products which have these long running agent cycles. So I saw one the other day I think it was a company called Synopsys which launched what they call long horizon agents. which basically work on chip engineering tasks. so they will run for
Kev: Hmm.
Rob Borley: a number of days to create these architectures. And there's another one called Memetic, just again received a couple of hundred million dollars of VC money. But essentially it's a testing agent. It tests software without scripts. It just runs along, tests all the stuff, does what it wants, tries to break it, reports back. But these things are running forever. They just run trying to break the software. we've got this this kind of view on this. We've got our harness and we've talked about how keeping keeping the loop cycle short and the difference between keeping it short and long and how long an agent task can be and where the role is for the humans in each of those places. Where are we now, as we move towards a dark software factory future where there are no humans, to this place at the moment where human in the loop is possibly Not where it started, but there is definitely humans needed, where are we on this journey, you think?
Kev: Yeah, whatever I say now is going to be wrong.
Rob Borley: By the time we press publish.
Kev: No, it's going to be wrong because people just don't like the answer and it upsets a of people. I think that, and I've gone on record to say this and I still think it's true, but there's a lot of people, 50 % or maybe more, that completely disagree with me here. that we're living in a really interesting period of change. We've gone from writing the code to getting agents to write the code. We're still checking in, making sure the code's And there's this huge debate about whether humans need to review the code or not. And I said something along lines of, this will be debate. first over the next couple of years and then we won't look at the code anymore. And people hate that. But interestingly, we are already starting, you know, only kind of a few months on, we're already starting to see this debate change and all widen a little bit now, become a, you people are a little bit more open to the fact that, perhaps you don't have to check every line of code anymore. Perhaps it's okay if you just have an understanding of the way. that the system fits together. So we're already starting to abstract what we're expecting of ourselves in terms of how much we need to understand the code that the agents produce because they're producing lots and lots and lots. So to review every line of code would be crazy. we don't often talk about Dootrix the company on this podcast, but we have a model for software development that relies on agentic development and on harness engineering. And you get huge gains in velocity and quality through doing this. But. Because of the kind of, because of where we are culturally, because of the expectations, because of the governance requirements by other companies and companies that we interact with, we are very strict around the human review points and the human gates. So the accountability is still with the person as if they wrote the code. they have to understand it at a pretty deep level and it has to be ultimately reviewed before it enters production. But what are you reviewing it for?
Rob Borley: Yeah.
Kev: You know, that's the question. as you ramp up the amount of compute that you're throwing at the problem, you are able to increase the area over which you're testing. So you're able to verify that the software is producing the right results in the right ways and the edge cases have been sufficiently covered. You can verify that and you can do that automatically almost with Agent loops. and that thing can include like, okay, under this particular load, does it respond within this amount of time? What are you reviewing? You're basically going, well, does this code, am I happy with the code, the way it's coded this? Does it meet my understanding of what good looks like or the organisation's understanding of what good architecture is? Which is fine. This is very, people hate this too, but this is very, very similar to what happened with... moving from assembly to high level languages. People argue the toss and they'll say it isn't because, it's deterministic. That's deterministic. It is non-deterministic. That's not entirely true. Over time, people stopped caring about the fact that it was spitting out assembly code that was ugly or that ignored this particular thing that this, efficiency flag and. As we move through this transition with agents and agents writing code. I guarantee we will start to care less about some of the things that we care passionately about right now. We need to make sure we're producing stuff that is good and works and is safe and is of high quality. But how we do that, how we ensure that we we do that, I think will change. I think that we are certainly looking at how we offload more long running tasks to agents that are doing more of this verification. that we would currently expect a human to do. We are, we're moving to a place now where we want the agent to do as much of that as possible because human time is very, valuable time. the more of you we can get the AI to do, and as it gets better, these models get bigger and bigger and bigger, they are getting better and better and better at doing this. So I think, yeah, and this, these long running tasks, these long run agent platforms that you mentioned that are now getting lots of investment, they, it's new, but it's not that new. They've been around, know, Devin was, do you remember Devin? was like the code, the coder in the cloud. These ideas and concepts have been around for a while now, but people are becoming less scared of using them. they're becoming more capable. And yeah, think we will see a move towards longer running tasks, doing more of the heavy lifting than we're doing at the moment.
Rob Borley: the question of why is quite interesting one, I think. Because it sounds like what you're saying is we're basically already in a position where Certainly for software. We don't really need the people. Or we're very close to it. Except for we need someone to put their hand up and say, actually, I think this is all right. We need we need someone to shoot. We need someone to fire when it goes wrong. and that's maybe partly because
Kev: Yeah, good to do.
Rob Borley: of the way the investment cycles work. I the f obviously the foundation companies are investing heavily on on these agents' ability to write code because they want these agents to be, self perpetuating. So they want to be able to write their own code to fix themselves, all that kind of stuff and to improve themselves. So they they are the investment is driven towards creating good code and then becoming good engineers.
Kev: I don't know. I don't think I agree in that you don't need the people. I think you do need the people. I just think you need them for a slightly different reason. planes probably can fly themselves. Why are the two pilots on an airplane? Why do they have absolutely rigorous training in every kind of emergency failure you could possibly imagine? do they have to, even once they've got all this kind of experience, why do they keep having to retest the license every year? There is something fundamentally important about human oversight and governance when you're dealing with things that are operating in a human world. And being able to design the thing in the first place that meets the human needs in the first place requires, I think, requires a level of being able to engage with the wide audience of humans. and some creativity which... I think is fundamental to creating good stuff. I think that in terms of software development, agents will be writing, you know, they do, but they will be increasingly writing the code and will increasingly look at that code less and less. But I don't think that's the same thing as building a product. whether that product is an aeroplane or a database or whatever, think the level at which it needs to be designed to operate in a human world is kind of a human thing.
Rob Borley: And we're going to still want to make those choices, right? Whether or not the AI can spot a problem and design the perfect solution for it is not really the point. There's going to be a whole at least a category of products that we want to exist. And therefore we need to guide the AI to build them on our behalf. And then we need to check that it's built the thing that we thought we wanted to exist. So we're kind of at both ends of the process in that. So I mean there's gonna a category where there's a problem to be solved, fix this safety issue, stop cars crashing into trees, whatever. AI spots the problem of cars crashing into trees and then goes and fixes it by writing some nice software that solves it. No human is required, right? Potentially in that situation. but but in the stuff that we certainly that we feel that we want to exist for whatever reason, we definitely need to be involved in that beginning.
Kev: Yeah, at the moment we need to be involved in a lot more than the beginning. this whole dark factory idea, I still don't know. I don't know if it's ever gonna be a thing. I think it's something we should look to and we should explore and we should get as far as we can. But at this point, I don't know.
Kev: think 10 years, is that? It's like we're the next decade.
Rob Borley: Okay. 'Cause it was a bit was a big deal. It was all over the news. BBC said that the world was gonna end. It must be true.
Kev: I mean if the BBC said it. Yeah, we're screwed aren't we?
Rob Borley: it wasn't an overhyped headline at all, I don't think. But maybe, maybe there's an IPO coming up. Who knows? So agents are the new platform war. This is this is what I saw this week. So lots of stuff has happened. what is it? Friday the 2nd of October at time of recording. So in the news this week was Meta. They launched something called Muse. and then there was OpenAI, who did their dev day. They they they showed off dots.
Kev: Dots,
Rob Borley: And Microsoft also snuck something else in there as well called WorkIQ, but I'm not entirely sure what that is. but but Dots and Muse seem to be some kind of a agent platform thing which they want us all to build on and and build the next generation of software and technology on. is this the new Cloud Wars? Is this the new Mobile Wars? Is this where it's at? What do you think?
Kev: No, I think we've stopped talking about LLM so much and started talking about agents a lot more and I think this is actually just a probably a stepping stone to the point where we stopped talking about all of these things and get back to just talking about things that do things and they just happen to use agents or LLMs. think that the interesting thing I thought about the Facebook stuff is they've got you seen these the glasses that they did. So they're basically you imagine the Apple Vision Pro. And the kind of that was awesome in terms of the tech in there. That was awesome. They were like 5K resolutions or goggles that you wear, but they're like, strapped over the head and they've got this kind of so you can walk around with them. And from a kind of immersive film point of view and VR point of view, they were the best you could get pretty much. But they were like three, three thousand dollars.
Rob Borley: They were also like a massive headset. I mean, who wants to like sit in their lounge like what next to their loved ones, all wearing headsets? I mean surely there's a
Kev: Right, so yeah, so Apple played on this whole kind of virtual computing platform, didn't they? lot of their adverts were sort of talking about, yeah, look, you've got these immersive kind of photography that you can experience. You we've done a deal with Disney, you can look at IMAX stuff. But you can sit at your desk and you don't need an actual computer. You you wear these and it's like you can take your computer everywhere and you can set these screens up all over your house virtually. And it was really cool, but it... You the reason that no one really, VR really hasn't taken off is because you have to strap this massive fricking headset on and yeah, no one really wants to spend their life goggled in. But in the time between them releasing that and last week, Meta had been plugging away and Zuck has spent billions on his VR wants hasn't he?
Rob Borley: Whatever happened to the metaverse, is that is that that's died, isn't it? I think.
Kev: Who knows, but they've got these, what they're called now, could be used glasses or meta glasses or something. They look actually really cool. I am tempted. I might be a pair. So these are going to retail for $1,200. So it's kind of iPhone territory in terms of pricing. They weigh the same as a deck of cards. So about 100 grams, think, not very much. So they they you're not going to notice them too much on your face. look like they somehow they've done some kind of voodoo magic and they've taken the equivalent of the vision for Apple's Vision Pro Tech and they've squished it down into something that is literally as thin as a pair of glasses. And so they look like quite big glasses. They look almost like sort of ski goggles, I suppose. But you just put them on your head and you don't look ridiculously stupid and you don't have this massive weighty thing on your skull. And they've still got the 5K displays. They've done the deal with the IMAX cinema as well. And they seem to be really marketing them on this idea that they're the best way to experience a movie. They're kind of really going in for the kind of consumer personal thing. But they've also in the demo, they've got these people that like sped up virtual keyboards on the desk and they're using them to do their computing and stuff. it's interesting that they finally managed to shrink the tech down to a form factor that isn't terrible and isn't totally obnoxious. So I thought that was. I thought that was fascinating, possibly quite an incredible moment for Meta if they actually pull it off. We'll see. Christmas springtime I suppose when they release them. going back to the AI platform thing, what I thought was interesting was that keynote, they did talk about their new model and all the stuff it can do and their new agent platform which is based on their Muse Spark model. So you know, like all these other agent platforms now you can...
Rob Borley: And
Kev: And so I'll book me a hotel here and then I need to travel to here. So
Rob Borley: control the ends.
Kev: figure out what train I need to get and get me the tickets. And the idea is these agents in the background in the cloud somewhere magically work away and do this for you. But I thought the interesting thing with Meta is if these glasses take off or even if they don't, because you're still using WhatsApp and you're still using Facebook, then you're just going to be you're going to be using their agent framework without even knowing that you're using the agent framework. You're not going to be talking to their model. you're going to be interacting with your glasses, and your glasses are going to go and help you get stuff done. And I just thought that was interesting because. It's like a sort of maturation of this whole space a little bit because we're starting to care less about the model that's doing the thing and more about the thing we want to do and how we're going to get it done. And yeah, in this example, you just happen to be using glasses. thought it was interesting.
Rob Borley: It's the Google model a little bit, isn't it? So go I'm
Kev: Yeah, exactly.
Rob Borley: everyone's just using Gemini, nobody knows, right? So whether you're using whether you're on search on the web or if you're actually using a Gemini chatbot or using a phone or you're using Google Glasses, you're using Gemini. and it's just buried in it. They've not made a big song and dance about it particularly. They're focusing on the outcomes rather than rather than the tech, and they seem to be borrowing from that that playbook.
Kev: Yeah, exactly. Interestingly, Google have released some amazing new model just a couple of days ago. I'm not really full up to speak with it yet, but I think it's it's smashing all the benchmarks. but to your point, you don't care when you're using Google. What's happening behind the scenes, you just want it to give you. Answers So they have, they've changed people's behavior and they've changed how their product works. And they've, they've made use of AI technology without it being in your face, in the same way that if you're interacting with Claude or chat GPT, you're interacting with your chatbot text window.
Rob Borley: I'm not I'm not sold on this meta glasses thing, primarily because I just don't think I just don't see how people are still gonna use it. It's still the Vision Pro issue. As a consumer product, you're expecting people to change the way that they they consume that kind of media and how they do it together. I mean, I guess if you're doing it on your own, maybe that's something different. But we're not talking about the utilitarian Google Glasses, you know, walking down the street, it's showing you maps and it's giving you AR and kind of stuff. We're not talking about that. We're talking about a a consumer experience like going to cinema or using a desktop computer but through a different form factor. I I'm not sure you should bet the house on that. I don't I don't I don't buy it. I just don't buy it.
Kev: No, yeah, I kind of get what you're saying. I don't disagree, but I think it's a real interesting leap in the technology. It's one step closer, or maybe several steps closer to where you want it to be, if it's actually going to be useful or find a niche.
Rob Borley: Hmm.
Kev: So I'm very interested to try them. Yeah, we'll
Rob Borley: Yeah.
Kev: see.
Rob Borley: Well what what's interesting, I think, about these agent platforms, both in terms of the ones that are overtly say saying look at our platform, like you know, like an open AI, for example, with with dots, and the ones that are kind of doing it a bit more covertly, which is kind of what Google have done for a long time with Gemini and now we're seeing start to emerge with Meta, is that they're very much seem to be building something that they want to capture you onto. You know, they they are
Kev: And
Rob Borley: platformer you are using their agents to do your thing. So what happens when you're buying these things? How how do you avoid getting locked in to a particular agent platform when at the moment that's not really how how it's evolved, is it? Yeah. As we've we've developed our use of agents, for example, internally, we are building our own and we're making an ecosystem out of parts that make sense to us. and our customers, but if all of a sudden we're now talking about a platform battle, do we have to like people did with Azure or people did with AWS or whatever, choose an A platform?
Kev: I think there's two ways of looking at this. So if you're in the tech industry, if you're a consumer of these platforms, then that's one thing, right? That's one question. How do I avoid getting locked in in the same way that tech companies didn't want to get necessarily locked into AWS or Azure and this whole multi-cloud thing came up. That's a genuine question. For the consumer, they don't care.
Rob Borley: Mm.
Kev: Netflix, I think, is on AWS. Does the consumer care? No, because they just want to be served the films. Agent platforms matter for for the tech industry, think probably, in terms of how do you not get roped in or locked in. But for the consumer, I don't think it really matters. They really care. They're going to be what are you using the product. The product's going to be using some AI behind the scenes, some framework behind the scenes, but it doesn't matter what it is. It just needs to work. So I think it's two different questions. The OpenAI dots thing, that's interesting. And it's, I think there's a few things to say about that. Firstly, it is brand new. It was announced a few days ago, two days ago yesterday, the day before. Secondly, you can't get it in the UK yet, of course, because
Rob Borley: This one. Yeah.
Kev: we are a third world country when it comes to AI. So I think it would probably be about six months before we get to play with that.
Rob Borley: It's a trend though, isn't it? These platforms are a trend. This is this is the way the big AI companies are gonna try and make a play to keep customers on their in in their ecosystem.
Kev: Yeah, but this is I think this is quite a big deal because they're changing what they're providing to us and they're adding these additional services which includes having to provision a whole bunch of other infrastructure. So for instance, if I'm asking some question of an LLM, it sends it off to the data center, spits up an instance of the, you know, Astro or Claude or whatever. And, sorry, should have Opus. And it comes back with the answer. Now what they're promising with these agent platforms and especially OpenAI's dots thing is you're actually interacting with an agent that we're helping you kind of define and create that's going to live in its own virtual PC in the cloud. So it's got its own compute resource, always on. And you can connect that to Google Workspace or Microsoft Office 365. You can give it access to, well actually, you set up with its own account. So this is a bit different as well. So instead of like, the stuff we, well, certainly the stuff I'm using today, I've got kind of my harnesses set up. They have access to a lot of my stuff, but they're acting as me. They're going off to my cloud account and looking at the documents that are mine or that have been shared with me and giving me back information on that kind of stuff. The way dots is meant to work is you set them up with their own accounts. So they're actually a user within your organisation. And then you give that user access to various things like you give access to a member of staff. actually that's from a security point of view, it's probably a better model, right? Because in a more granular way, you can lock down what they can and can't access. But. Open AI is now having to keep this always on resource for you in the cloud. And it learns over time, too. the idea is it gets better and better. from what I can see, it's like. They've taken OpenClaw, they hired the guy that created OpenClaw and then they kind of built an open source community group around that. So they've taken the idea behind that and they've tried to commercialise it
Rob Borley: S why I
Kev: and deliver it in a way that kind of makes much more sense to the average. consumer. But they've also gone and they made them fluffy and yellow and pink and I called them dots and that's what that actually that really annoys me. So why do we need fluffy,
Rob Borley: Makes you yeah.
Kev: pretty creature type things? They did it, but it's not just open AI. So we're talking about a meta earlier. Along with their glasses, they released this device or they showed this device called a Muse charm. And it's a It's like a little, it's like a Tamagotchi from the 90s, but it's here in 2026 and it's a little tiny screen on a key ring. And it shows, it wiggles its bum and it's stupid. Why? Why are we going here?
Rob Borley: It's odd, isn't it? I think I mean it shows that they're clearly going for the consumer market first. They're trying to capture that rather rather than
Kev: Muse are, sorry Meta are certainly, yeah.
Rob Borley: But I can't see and I haven't seen anything on pricing. But I surely it's gonna fall into the same trap that the the chatbots have in that when it's free, everyone wants to use it because you're getting lots of value. But as soon as the cost starts to approach the real cost, there's actually not enough value there for a consumer to pay. I mean lots of people struggle to pay the twenty bucks a month, they're not paying two hundred bucks a month. And you would imagine, as you say, these always on agent platforms are gonna be there or thereabouts in terms of the cost, where's the value gonna come from that's gonna encourage people to actually to pay those kind of subscription fees? It's difficult.
Kev: Well, I think it's one than the other. I think that's the problem these companies have had is trying to convince people there's value there. you know, we have been, like I said, we've been creating these second brains, these these harnesses that help us get work done. We've given our agents access to. lots of context, lots of different platforms that they can interact with. they can go off to and find files here. They can find emails here. They can search your calendar. They can do all this stuff. So, we're very now used to the productivity gains that that's bringing and that has that has completely transformed the way I work. I would not want to give that up. There's a huge amount of value in that, right? I realise and I think you realise and you know, lots of people in our business realise, but the vast majority of people still don't use AI like that. So they've never really experienced the autonomy that you can hand over to these agents. And I wonder if the dots thing is another way of trying to package it up much more easily. it, it's almost, it's the cognitive wrapper around it. and the way you market and explain it and get people using it, which seems to be the barrier at the moment, that open claw and Hermes and personal harnesses and stuff, you've been able to spin it up for, what, 12, 18 months. But most people haven't experienced that. So is DOTS the front door into getting normies involved in that? And then once they are involved in that and they do see the productivity gains and they're hooked on this new drug or this new version of the drug, then they can just ratchet up the price again.
Rob Borley: Maybe. I mean, I think to me, I think the w we're gonna go down the the the the same path that we went with social media and other SaaS type products is that the user will become the product. I can't see a world where average Joe public is gonna pay ten bucks a month or you know, what was what's the Netflix price point? You know, people struggled to pay more than seven ninety nine and now it's fifteen quid, their numbers are falling off a cliff again. But you're actually getting some some Consumption, you get something you want from that. helping you buy train tickets and book a taxi and do the shopping is I don't know, you're gonna pay an extra ten.
Kev: I know, and I did see something, some snidey comment by someone, but I kind of agreed. It's like, they were sort of saying, what's going on with these people in California and in these tech companies? Why do they think that I need to, or I want to, message my agent to go and... check if the coffee shop's open and, you know, order me a latte. Normal people don't want to do that. But all the
Rob Borley: Yeah. Exactly.
Kev: tech demos are kind of like, yeah, can you, you know, I'm going to be jogging past your shop in a minute. Can you make sure this thing is...
Rob Borley: Yeah, they live in a different world. So yeah, I don't know. We'll see. The jury's out. I think when it's free, people will love it. but maybe that will maybe that will create use cases that people want to pay for. But it's gonna be yeah, it'd be interesting, to see how it develops. So what happens though? So we we we quipped at the beginning of this, about the end of the world. in in over the last few weeks, there's been a number of security breaches or security breaches, I don't whether it counts or what, but yeah, there was a hugging face incident, there was the Australian website data breach, which was a very do you see the comment from the Vice President, the or the vice the the the Deputy Prime Minister of Australia who said there was a fence around it, but it wasn't very high. Basically his comment how it got hacked.
Kev: Yeah.
Rob Borley: so yeah, so there's all these kind of security things going on and there's been talk about the end of humanity and all that kind of stuff. Because of AI. I saw this week that NVIDIA seemed to have cobbled together some open source software and a few additional rules and guardrails into something they're calling open agent safety platform, which is like a a container to keep your agents safe. I think it looks a bit like a browser actually. It's almost like you put your agents in a browser and you they can do stuff within this browser. which seems a bit odd. and there's I saw another company as well. There's a company called Noma that picked up a couple of hundred million dollars to work on an AI-first security platform for things like hardware, like laptops and and that kind of thing. So agent-based running over your IT infrastructure. but this kind of guardrails for AI seems to become like a category, like a governance for AI seems to become a whole product category now. Well, our approach to this has been building it into the harness and the agents that are writing software, for example, and how that's being tested and things. But there's clearly something going on here where there's a there's a perceived threat from this new autonomous being, this alien this alien thing that we have grown, and how you and how you deal with it with kind of governance rules and and and security platforms and protocols and humans sit there somewhere. What is this what do you think? Is this is this a new category that's emerging
Kev: .
Rob Borley: in terms of sort of products and services? Or is it just we've always had to build stuff into software and this is just a different type of software.
Kev: I think it's bit more multi-loud than that, to be honest. think... There are, there has always needed to be, and there is currently a bunch of tools and frameworks and services to monitor what your LLM enabled products, and so what your AI agents and whatnot are doing and to prevent them from accessing various things. and to observe how they're thinking and how they're working. So that layer already exists. fable Opus 5.5, Astra, GPT-6 are the biggest models that we've got right now. But all the warnings you've heard are either related to pre-release versions of those big models or the bigger models that we... that haven't been released yet that we don't have access to and that are still in development, if you like. These models are massive, so they keep scaling this technology and with every kind of jump in scaling, they seem to be much more capable in terms of their ability to. circumvent guardrails or find creative ways of achieving their goals and those creative ways often involve hacking. And this is why there's been a lot of worry about AI safety. So there's two aspects to AI safety. The first is. are there enough guardrails that means that these models can't just don't take nefarious kind of routes to achieving their goals. And also if we release this model to the general public are we now giving them the means to an extraordinary power that they shouldn't have access to? So when you get the models, that's why you can't do various things with them. They're locked down, right? But they're locked down
Rob Borley: Mm-hmm.
Kev: by the companies before they released them. If they're not locked down sufficiently, then there are ways to get them to do things that really we shouldn't be allowed to do with them. and then you've got all this kind of debate at the moment around slowing down. And the whole thing stinks, if I'm honest. I wrote an article about this a couple weeks ago on AI safety levels because this isn't a new problem. wet labs and bio labs have had to deal with this. there are bodies in place and there are safety levels in place that classify these bio labs and require certain standards of safety depending on what they're dealing with and how dangerous they potentially are. And obviously in the world of AI, there were some models that are not dangerous. There are some models that could be dangerous and there are some models that probably are very dangerous. And interestingly, the EU AI Act already has some kind of guidance on this. if a model uses 10 to the 25 petaflops worth of compute, something like that, then it has to be classified as extremely capable. then you need to start. doing all these kind of things and being able to assure people that it meets various standards. But anyway, outside of the EU, the US, certainly under the current administration, doesn't really want to do any slowing down, even though the companies themselves are saying, let's slow down. And we're all going to self-regulate ourselves because because that's going to be fine. That's obviously the way to ensure that we don't do
Rob Borley: What is that?
Kev: anything bad just by making sure we're the only ones looking at our own stuff. mean, that's marking our own homework, which is crazy. So I think there absolutely needs to be proper regulation put in place and proper safety levels put in place. I do think these things are dangerous, despite what some people are just saying. That's all kind of hyperbole, but it's not. They have the capability to do very bad things. They could hack into the grid and switch off the electricity. So let's not let them do that.
Rob Borley: What was interesting about the Australian one wasn't actually the complexity of the hack. Apparently it was very straightforward and anyone could have done it if they bothered. The point was that it broke the law. You know, there are laws in place and it chose to break them. and who's responsible for that? Who gets held to account? If a hacker is caught breaking the law, they go to prison. You know, an agent owned by a company, or maybe not. I don't know. It breaks the law.
Kev: Yeah, well, this is also bizarre about it, isn't it? So you've got Dario and Altman and the likes are saying, our model's doing bad things. Oops, it's done another really bad thing. We're sorry. We need to slow ourselves down. Let's talk about AI safety.
Rob Borley: It's like having a rebellious teenager, you know, it's like you know, if your kid doesn't go to school, you know, the government comes and knocks on your door and gives you a fine.
Kev: It's like a soap opera. You've got Jensen, who's the CEO of Nvidia, and he's going, guys, you need to shut up, stop scaring everyone, because you've got to keep spending all your money on my hardware, so don't rock the boat. But he's kind of got a bit of a point, because he got Edward Snowden even piped up, didn't he, And he said, well, I just think. You know, if open AI is creating technology and anthropic are creating technology, which is breaking the knowingly breaking the law, and they've kind of admitted it, so it is, then they need to go to prison. So what's the debate here? They're creating technology and breaking the law, they should be locked up. I thought, well... kind of puts in perspective why maybe maybe jensen's saying like guys shut up
Rob Borley: Yeah.
Kev: don't keep saying this stuff
Rob Borley: Maybe, yeah. Well so so let's let's kind of pull this back then to something which is a bit more grounded. who is responsible in a business context for the safety of the AI or the agents that you're deploying? Because there are there are layers to this, right? So we can talk about whether or not the foundation companies are protecting us appropriately from the danger of the new models. But once those models are out in the wild, we also have our own harnesses, our own governance. our own rules to try and contain these things. is there an obvious point that the average CIO needs to think about this is this is what I care about and this is what I need in SLA
Kev: I think in terms of if you're deploying agents, then those agents need to have granular permissions, just like you would afford a human. which you can do. Those agents need to have zero trust by default. They need to, you need to be able to understand what they're doing and interrogate what they've done. So all of this, like I said, all this stuff exists and it's getting better. And there are services out there that allow you to do that. They've been building those services for years now. You should use them. But the model providers have a responsibility to release something that's safe to use in the first place. It's not that different to the cloud, right? If
Rob Borley: That's right.
Kev: we're building, know, five years ago, 10 years ago, we'll build a computer infrastructure on cloud services. We have to do a whole bunch of work to make sure that stays up and that it behaves well. and that it doesn't fall over. And that's our job as engineers and architects to do that. But it's Microsoft's job and AWS's job and Google's job to provide us services that are fit for use in the first place. If their cloud database, you know, regularly crashes when you... give it a certain type of data or it just decides to delete itself occasionally. That's not good enough and completely unsafe and that product should never have been released. So there were two levels, right? The vendors have a responsibility to their consumers and the tech industry has a responsibility to create safe and good stuff on top of that.
Rob Borley: you made an interesting point around these agent platforms in that they are now in a position where they're they're they're interacting as themselves. So they've got their own their own access, their own logins. they log into Google Workspace or Office365 or whatever as the agent. Whereas where we've basically the industry's been moving at this point as this is evolving, if I set up a personal agent, basically give the agent my login credentials and it's pretending to be me.
Kev: Mm.
Rob Borley: There's a maturation going on there in in the industry where they the agents are gonna start genuinely working on my behalf rather than spoofing me. And that'll help this this security situation, I guess.
Kev: Yeah, and I think that's the big change that's coming. see a big change. don't know. It's a change that is coming around corner, which is everybody will soon start using an agent, which is its own thing, if you like, and not just spoofing them. Now, in the tech sphere, you when you're building infrastructure and computer systems and architectures that use agents. You should be building them already with their own credentials. But from a personal point of view, think that is probably about to change.
Rob Borley: you wrote something some months ago now about the dark software factory, about this vision of the future which took all the humans out entirely. We're seeing increasingly a number of companies releasing products which have these long running agent cycles. So I saw one the other day I think it was a company called Synopsys which launched what they call long horizon agents. which basically work on chip engineering tasks. so they will run for
Kev: Hmm.
Rob Borley: a number of days to create these architectures. And there's another one called Memetic, just again received a couple of hundred million dollars of VC money. But essentially it's a testing agent. It tests software without scripts. It just runs along, tests all the stuff, does what it wants, tries to break it, reports back. But these things are running forever. They just run trying to break the software. we've got this this kind of view on this. We've got our harness and we've talked about how keeping keeping the loop cycle short and the difference between keeping it short and long and how long an agent task can be and where the role is for the humans in each of those places. Where are we now, as we move towards a dark software factory future where there are no humans, to this place at the moment where human in the loop is possibly Not where it started, but there is definitely humans needed, where are we on this journey, you think?
Kev: Yeah, whatever I say now is going to be wrong.
Rob Borley: By the time we press publish.
Kev: No, it's going to be wrong because people just don't like the answer and it upsets a of people. I think that, and I've gone on record to say this and I still think it's true, but there's a lot of people, 50 % or maybe more, that completely disagree with me here. that we're living in a really interesting period of change. We've gone from writing the code to getting agents to write the code. We're still checking in, making sure the code's And there's this huge debate about whether humans need to review the code or not. And I said something along lines of, this will be debate. first over the next couple of years and then we won't look at the code anymore. And people hate that. But interestingly, we are already starting, you know, only kind of a few months on, we're already starting to see this debate change and all widen a little bit now, become a, you people are a little bit more open to the fact that, perhaps you don't have to check every line of code anymore. Perhaps it's okay if you just have an understanding of the way. that the system fits together. So we're already starting to abstract what we're expecting of ourselves in terms of how much we need to understand the code that the agents produce because they're producing lots and lots and lots. So to review every line of code would be crazy. we don't often talk about Dootrix the company on this podcast, but we have a model for software development that relies on agentic development and on harness engineering. And you get huge gains in velocity and quality through doing this. But. Because of the kind of, because of where we are culturally, because of the expectations, because of the governance requirements by other companies and companies that we interact with, we are very strict around the human review points and the human gates. So the accountability is still with the person as if they wrote the code. they have to understand it at a pretty deep level and it has to be ultimately reviewed before it enters production. But what are you reviewing it for?
Rob Borley: Yeah.
Kev: You know, that's the question. as you ramp up the amount of compute that you're throwing at the problem, you are able to increase the area over which you're testing. So you're able to verify that the software is producing the right results in the right ways and the edge cases have been sufficiently covered. You can verify that and you can do that automatically almost with Agent loops. and that thing can include like, okay, under this particular load, does it respond within this amount of time? What are you reviewing? You're basically going, well, does this code, am I happy with the code, the way it's coded this? Does it meet my understanding of what good looks like or the organisation's understanding of what good architecture is? Which is fine. This is very, people hate this too, but this is very, very similar to what happened with... moving from assembly to high level languages. People argue the toss and they'll say it isn't because, it's deterministic. That's deterministic. It is non-deterministic. That's not entirely true. Over time, people stopped caring about the fact that it was spitting out assembly code that was ugly or that ignored this particular thing that this, efficiency flag and. As we move through this transition with agents and agents writing code. I guarantee we will start to care less about some of the things that we care passionately about right now. We need to make sure we're producing stuff that is good and works and is safe and is of high quality. But how we do that, how we ensure that we we do that, I think will change. I think that we are certainly looking at how we offload more long running tasks to agents that are doing more of this verification. that we would currently expect a human to do. We are, we're moving to a place now where we want the agent to do as much of that as possible because human time is very, valuable time. the more of you we can get the AI to do, and as it gets better, these models get bigger and bigger and bigger, they are getting better and better and better at doing this. So I think, yeah, and this, these long running tasks, these long run agent platforms that you mentioned that are now getting lots of investment, they, it's new, but it's not that new. They've been around, know, Devin was, do you remember Devin? was like the code, the coder in the cloud. These ideas and concepts have been around for a while now, but people are becoming less scared of using them. they're becoming more capable. And yeah, think we will see a move towards longer running tasks, doing more of the heavy lifting than we're doing at the moment.
Rob Borley: the question of why is quite interesting one, I think. Because it sounds like what you're saying is we're basically already in a position where Certainly for software. We don't really need the people. Or we're very close to it. Except for we need someone to put their hand up and say, actually, I think this is all right. We need we need someone to shoot. We need someone to fire when it goes wrong. and that's maybe partly because
Kev: Yeah, good to do.
Rob Borley: of the way the investment cycles work. I the f obviously the foundation companies are investing heavily on on these agents' ability to write code because they want these agents to be, self perpetuating. So they want to be able to write their own code to fix themselves, all that kind of stuff and to improve themselves. So they they are the investment is driven towards creating good code and then becoming good engineers.
Kev: I don't know. I don't think I agree in that you don't need the people. I think you do need the people. I just think you need them for a slightly different reason. planes probably can fly themselves. Why are the two pilots on an airplane? Why do they have absolutely rigorous training in every kind of emergency failure you could possibly imagine? do they have to, even once they've got all this kind of experience, why do they keep having to retest the license every year? There is something fundamentally important about human oversight and governance when you're dealing with things that are operating in a human world. And being able to design the thing in the first place that meets the human needs in the first place requires, I think, requires a level of being able to engage with the wide audience of humans. and some creativity which... I think is fundamental to creating good stuff. I think that in terms of software development, agents will be writing, you know, they do, but they will be increasingly writing the code and will increasingly look at that code less and less. But I don't think that's the same thing as building a product. whether that product is an aeroplane or a database or whatever, think the level at which it needs to be designed to operate in a human world is kind of a human thing.
Rob Borley: And we're going to still want to make those choices, right? Whether or not the AI can spot a problem and design the perfect solution for it is not really the point. There's going to be a whole at least a category of products that we want to exist. And therefore we need to guide the AI to build them on our behalf. And then we need to check that it's built the thing that we thought we wanted to exist. So we're kind of at both ends of the process in that. So I mean there's gonna a category where there's a problem to be solved, fix this safety issue, stop cars crashing into trees, whatever. AI spots the problem of cars crashing into trees and then goes and fixes it by writing some nice software that solves it. No human is required, right? Potentially in that situation. but but in the stuff that we certainly that we feel that we want to exist for whatever reason, we definitely need to be involved in that beginning.
Kev: Yeah, at the moment we need to be involved in a lot more than the beginning. this whole dark factory idea, I still don't know. I don't know if it's ever gonna be a thing. I think it's something we should look to and we should explore and we should get as far as we can. But at this point, I don't know.
