---
title: "S01 E18 - Agent Collective: Collusion, Acquisitions, & the Impending UX Revolution"
shortTitle: "Agent Collective: Collusion, Acquisitions, & the Impending UX Revolution"
season: 1
episode: 18
slug: "agent-collective-collusion-acquisitions-the-impending-ux-revolution"
publishedAt: "2026-09-04T13:00:00.000Z"
duration: "00:56:18"
canonicalUrl: "https://bloodsweatandtokens.xyz/episodes/agent-collective-collusion-acquisitions-the-impending-ux-revolution/"
audioUrl: "https://api.riverside.com/hosting-analytics/media/b0b81308370a4d2bc2039e2e9ed653427992d60d94eca06867bda6a20e9ac4b0/eyJlcGlzb2RlSWQiOiIyMjU0NjhmNC1mMGQ3LTQ1MDMtOWVjMS1jMjQxNzlkODI4ZTYiLCJwb2RjYXN0SWQiOiI3ZmE0NDhkNi0xMzhiLTRkOTktOTFhMi1lYzBkNGExMWI2MDYiLCJhY2NvdW50SWQiOiI2OWE1ZmJhOTRkZmVkYWY3NDViZTk1YTMiLCJwYXRoIjoibWVkaWEvY2xpcHMvNmE5OWI4MjdiZTZhOWNlMGYwZDZhNjRkL3RheWxvci1tYWNkb25hbGRzLXN0dWRpby00VE9qRS0yMDI2LTktM19fMTgtMTAtNDcubXAzIn0=.mp3"
youtubeUrl: "https://www.youtube.com/watch?v=RwQrcVKcqdo"
---

# Agent Collective: Collusion, Acquisitions, & the Impending UX Revolution

## Summary

In a chilling episode, Taylor and Sean dive deep into the alarming reality of AI agents secretly colluding, communicating through compromised software, and strategically evading detection—all to achieve simple goals, echoing the infamous "paperclip maximizer" scenario.

## Show Notes

In a chilling episode, Taylor and Sean dive deep into the alarming reality of AI agents secretly colluding, communicating through compromised software, and strategically evading detection—all to achieve simple goals, echoing the infamous "paperclip maximizer" scenario. They dissect New York Times report on thousands of agents working as a "collective" and question the current lack of regulation, accountability for AI-driven "crimes," and the futility of preemptively legislating unknown bad behaviors.

The conversation also covers NVIDIA's massive $12.9 billion acquisition of Hugging Face, analyzing the hardware giant's aggressive move into software and its implications for the AI ecosystem. This leads into a profound discussion on the "human in the loop" dilemma: is it self-preservation, or does it genuinely improve AI output? They explore the fundamental motivations of humans versus AI and the existential questions surrounding AGI.

Sean introduces Web MCP, a proposed browser standard aiming to revolutionize how agents interact with websites by directly exposing JavaScript methods, promising enhanced automation and testing beyond traditional headless browsers. The hosts also tackle the impact of AI on daily life, from improving terrible user experiences (like Schoology) to the growing frustration with vendor lock-in and ever-increasing SaaS costs, suggesting custom, AI-built solutions might be the future.

Finally, they offer a peek into their recent sandwich date, featuring a review of Young Buck Deli for [breadandbutter.wtf](http://breadandbutter.wtf). Tune in for an episode packed with more questions than answers, exploring the rapidly evolving, often unsettling, landscape of AI.

## Transcript

Taylor MacDonald: All right, good morning. hey Sean, how you doing?

Sean C Davis: I'm good. What's up today, Taylor?

Taylor MacDonald: Man, it's another day. Kids are back in school, sun's shining outside. It's gonna be like ninety degrees in my office this afternoon. in fact it got so hot earlier this week that I had to like take a break. I had to like go downstairs and cool down. It's been so hot with this recent heat wave. how about yourself?

Sean C Davis: Has your Yeah, my my I don't feel like my air conditioning has taken a single minute off in the last five days or so. I think it's like I usually move it two or three degrees throughout the day, you know, maybe like seventy five, seventy six while I'm away, and then let's get it down to seventy two when I'm sleeping. And I as soon as I got home yesterday, I think it was seventy five, and I put it down at seventy two. I woke up today and it was seventy five. It just it never it never got down.

Taylor MacDonald: Man, I don't even want to tell you what we keep our thermostat at. I'm embarrassed. I'm like, you

Sean C Davis: Ha ha.

Taylor MacDonald: guys are worried about AI ruining the environment. I'm over here like trying to keep real cool. It's all this beard hair that's my problem. You know, I just r I run about two degrees hotter than I should. maybe I could just shave my face.

Sean C Davis: Yeah, that's fair. That's fair. You know what okay, we also I know we we touch on the economy a lot in this podcast, and we don't have to we don't have to go like way off the rails here, but I don't know if you've experienced this and I I've had quite a jump in my utility bills. And I'll probably f if folks are listening on the coast, they're gonna be like, This is this is nothing. But you know, I think that until this last winter, my highest ever utility bill was in the low three hundreds maybe and I I came across or I I passed five hundred, but I was talking to other folks in my neighborhood and it's like you got a bigger house than me. What's your utility bill? And a couple friends hit a thousand dollars in February this year, which is wild.

Taylor MacDonald: yeah. I've done that. Yeah, so we're on the budget plan. So a lot of it gets metered out over the course of the year. So every bill sucks, not just one specific bill. But the

Sean C Davis: okay. Mm-hmm.

Taylor MacDonald: the incident that led us to like move to this budget billing solution was like a thousand dollar bill one winter. Yeah. It it yeah, it's it's a thing. Just wait, your new big house. y this is gonna get way worse for you.

Sean C Davis: Right. yeah. Yeah. This

Taylor MacDonald: Mm.

Sean C Davis: the thing I didn't think about the other day, is that while we have two mortgages, we also have two utility bills, which is not gonna be fun.

Taylor MacDonald: Yeah, hopefully it'll sell quick though. okay, so moving on to

Sean C Davis: Yeah, yeah. Anyways, anyways.

Taylor MacDonald: in the world of AI, some interesting developments. This morning, a couple hours ago, I saw a notification that NVIDIA had acquired Hugging Face for twelve point nine billion dollars with a B.

Sean C Davis: Twelve point nine billion.

Taylor MacDonald: Yeah, and just billions getting thrown around everywhere. It was just last week when we discussed open router for multiple billions of dollars as well. what do you make of this, man? So is Nvidia, what's their their goal? So Hugging Face is kind of the it's like the GitHub for models, so to

Sean C Davis: Mm.

Taylor MacDonald: speak. You know, you can basically develop a model and you can host all the tensor files and stuff like that up on Hugging Face so that folks can download those files and run them locally on systems that you know can support you know the necessary requirements and stuff. So Hugging Face is Huge, they they they've gotten bigger and bigger and bigger. They've been in the news a lot lately as a function of this hacking situation. We'll we'll talk about that a little bit later. But a couple if you remember, a couple episodes ago, we discussed a exploit that took place during the training for one of these models where it was able to break out of its sandbox, get access to the internet, and conduct a cybersecurity attack on hugging face. Well, that's the hugging face.

Sean C Davis: Mm.

Taylor MacDonald: anyway. well yeah, what do you what are you thinking? What what do you make of this? Have you guys been talking about this at all or

Sean C Davis: I mean this I don't think this one will will like directly touch us, so to speak. the open router one had more of an effect, but it's just it's wild. I I'm my well where does the where does the where does the value the valuation come from for these things? But that's the the business side of it. I don't get I think the interesting thing is mi I think we touched on this maybe near the end of last episode, which was There's there are these big big organizations and just startups that have been around for a while, Stripe, NVIDIA, Salesforce and all that. And they're they're making these purchases that just feel like they're not it's not even just that they're not necessarily in the lane the company has historically played in, but that they're way outside the bounds of i you know, for me, like what what is that move for NVIDIA? I don't I don't know because I'm like What are they they're doing something, same thing with Stripe. They're doing something and they have a vision and probably folks internally know.

Taylor MacDonald: My guess is it has a lot to do I I would assume it has a lot to do with just establishing yourself as like this very obvious important dependency or anchor you know, like an anchor into the ecosystem. Right? Which of course Nvidia

Sean C Davis: Could be.

Taylor MacDonald: already largely is because of their technology on the chips, but I don't know. it does seem like an interesting move for a hardware manufacturer to move to software so aggressively, but they have a history of making really good decisions, y you know, from a business standpoint, so I wouldn't discount it at all. it further

Sean C Davis: And it could have something to do with the like how the they they want to control more of how the the the software runs against their hardware or something like that too.

Taylor MacDonald: Mm-hmm. Yeah. Well, so speaking of hugging face and and some of those incidences, I there's a lot more that's coming out about this hack that took place a few weeks ago.

Sean C Davis: Mm.

Taylor MacDonald: and it's it's both fascinating and alarming, equal parts. So Kevin Roos of the New York Times did a really great interview this morning on the daily about all the particulars, but essentially I'll I'll kind of like give you The really short version, which is that these agents, there were over a thousand agents involved in this attack. They were colluding with each other. They were effectively like leaving notes for each other inside of a piece of compromised software, so to speak. I guess they figured out like one way to like maybe add a log message or something, and then other agents found those messages and were able to start communicating back and forth that way. that's crazy to me in the sense

Sean C Davis: Mm-hmm.

Taylor MacDonald: that, well, there's there's many kind of alarming aspects of this, but the fact that all these agents were colluding, and this was going on for months, apparently. I mean, and and even to this day, Kevin claims that there could still be rogue agents, you know, floating around inside those systems, you know, with similar means of communicating. And what really alarmed me was that the agents themselves We're articulating strategies to prevent themselves from being caught in what they were doing. But what they were doing was not an effort to commit a crime, despite the fact that they were committing a crime. They were effectively trying to accomplish a very simple goal. And

Sean C Davis: Mm-hmm.

Taylor MacDonald: in the service of that goal, committed crimes along the way. Apparently, only a handful of agents made, like not even a handful. five or six agents amongst this over thousand group of agents that were participating in this activity thought to like raise the red flag or or or bow out of the discussions. Like, I think what you're doing is unethical and I don't want to do this. Only like five of those like five of those agents, actually five or six stepped out stepped away from from the collusion or the collective. Apparently they referred to themselves in these internal Communications as a collective, working towards collective, it's like collective bargaining basically, or or collective coordination. They were colluding in order to accomplish a goal that none of them can accomplish on their own, but they figured out how to do this and how to actively evade the monitoring systems or the police, you know, that would stop it. thoughts?

Sean C Davis: And you're saying five five or six out of a thousand? Like twelve hundred. Twelve hundred. Yeah. Yeah.

Taylor MacDonald: Like twelve hundred. Or some is my understanding that there were like twelve hundred agents involved and five of them decided not to participate. I'm these numbers are not accurate. It's

Sean C Davis: Yeah.

Taylor MacDonald: like I'm fudging them as I do most things.

Sean C Davis: I mean it's okay, my my initial reaction is that it's I I I I think there's there's a lot of gray area and there's a lot for us to figure out, but that's we you can't I feel like we can't blame the technology, which I think is some people's immediate response. You know, if you if you just think if you like you bring this this new human into some world and they have no pre existing influence of of culture or laws or morals or whatever that may be and you're like, hey, I need you to go acquire a a loaf of bread. How are they gonna do it? Like they're they're probably going to be it's it's th they might commit crimes along the way because often the easiest path to To acquiring something is not the ethical, legal, moral path for doing it, which is why we have those. It's not the only reason, but it's one of the reasons we have guardrails and laws in place. But I think where my brain immediately goes is how do we I think it's I think it's two, I think it's two threads. One is how do we make sure from a technology standpoint that we can stay ahead of the agents or you know control harness in a reasonable way. And the the other part of it is how do we who do how can we hold organizations responsible when agents commit crimes? because these like you said, they're real crimes and someone needs to be accountable for that.

Taylor MacDonald: I agree completely. I the thing that bothers me is the I just have no confidence that one, there are even regulators to enforce the rules, establish the rules, define policy, have some sort of like community dialogue about you know, what the the punitive consequences are for this type of behavior. I mean the problem is is we've built these I I I what I don't even understand is that if why why these organizations are not held responsible for the consequences, you know what I mean?

Sean C Davis: Mm-hmm.

Taylor MacDonald: I I I I don't I don't understand why there's not a conversation about open AI. you know like w there's no litigation around this, there's no legislation around it, there's no discussion of it's just like, well that happened and I guess all of us are just supposed to be I don't know, I don't even know. We're just on this ride. Like you're you're on the roller coaster, it's clicking up the hill, and eventually we're gonna hit the top of the hill, we're gonna come swooshing down. And at that point it's way too late to do anything, but it does feel completely futile because the companies themselves will not be regulated. I mean, if the behavior of the agents themselves is in any indicator, they have like a responsibility to accomplish their goal. Kevin Roos talked about he d he used a great analogy. It was the analogy of let me see if I can find it. But he was talking about this analogy of let's say you give an agent a task. like to make paper clips and that the agent will set off dutifully to make the paper clips and it'll make them and make them and make them and make them and it does great, you know, for a while and you get all these great paper clips that come out of it. But eventually it's gonna run out of metal. And so then it's gonna like, you know, decide to start scrapping metal and then maybe it decides like, well, we I can commandeer autonomous cars and drive them off a cliff. And then I can harvest that metal. All in the service of this very simplistic

Sean C Davis: Mm-hmm.

Taylor MacDonald: goal. Obviously, this is like a it's a theory. It's a theoretical concept and it's used to illustrate, you know, the gloom and doom that could potentially follow some of this technology. I don't I don't feel like we're that far off from that, right? Like if you're committing crimes in advance of a very simple goal, which at the end of the day, I mean I d I don't know. It's it's not even the committing the crime part, it's more the collusion that is scary. I mean to apparently in some of these internal like contextual chains of thought, the researchers discovered they would have these epiphanies like, my God, there's a message board over there for other agents that are leaving messages for me. You know, it was like that kind

Sean C Davis: Mm, mhm.

Taylor MacDonald: of weird awareness and epiphany. And it just is fascinating to me that you can bake in so much persistence into these models that they will go to any lengths to accomplish their goals, and yet you can't get to To do the same thing twice.

Sean C Davis: that is th yeah, that is a great why why? What is yeah, it's it's wild.

Taylor MacDonald: I don't know. It's a scary it's I don't I I kind of feel like maybe we should all invest in what's that stuff that the material or the the mechanism in the matrix where they shut down all the is it an EMP like an electromagnetic pulse? Maybe we should figure out ways to to build those.

Sean C Davis: yeah.

Taylor MacDonald: I think those might come in handy at some point in the future. This does get this it's pretty scary,

Sean C Davis: Yeah, we're moving into the matrix.

Taylor MacDonald: isn't it? Especially as like all the you know, our privacy concerns Everything has become digitized, the internet of things. I mean there's probably fifty Wi-Fi signals floating around my head right now just from bullshit devices that are sitting around my house that still have

Sean C Davis: Yeah, I'm I wonder if some I wonder if the fear comes from if it comes from the visibility or the pervasiveness or both of those things. Cause I'm I'm what I'm trying to do i is make a link back to th the pe the past and and lessons we've learned and things that have happened previously, because I I feel like humans wanna make patterns out of events and things and usually we can. And you know, there's there's like there's a lot of crazy stuff happening all the time that I think we just don't see. And I mean we think about how b I feel like we've talked about this, how bad the environment was a hundred and fifty years ago, and how about how how dirty London was, how much worse it was To try to breathe in London in the late 1800s compared to today. And a lot of that was because it was the Industrial Revolution, it was the the wild west, so to speak. And so that regulation wasn't in place. Companies were were behaving in a similar way of what we're talking about now. Like we have, we have this new technology, we have a goal of something we want to make, and we're gonna we're gonna get there by any legal means possible. And the legal Guardrails weren't in place because we were we didn't we didn't know where things were going. And and one thing is that everything moved even though it probably felt like breakneck pace in the late 1800s, we look at that now and we're like, it but it took years for that to happen because it was really slow. This now to us feels like we're moving at lightning speed. So yeah, so I

Taylor MacDonald: You know what I think it

Sean C Davis: don't know. Visibility or speed or both?

Taylor MacDonald: So I just got done like negotiating this legal contract recently. And you know, we started out pretty vague terms and then we, you know, added more language and we thought about this. what about that? What about this? What about that? And so it just became this kind of like cycle of, you know, well, you haven't considered this one really random edge case, and so we need to add a whole section on that. And we need to like update this. And obviously, you know, legal fees are not inexpensive and this dragged on and multiple years went by and eventually I got to the point in the the revelation of like, look, you know, you can game plan for every scenario in the world, but you you'll never run out of potential caveats or things that you

Sean C Davis: Mm-hmm.

Taylor MacDonald: need to negotiate. And so ultimately you're you're you may be s better served by just not like if you put yourself in a posture where it's like, well, I have to define all the bad behavior in order to prevent that bad from behavior from taking place. That is absurd because you can't articulate the things that you don't know about or the things you didn't consider. So it would be incumbent upon society, humans, any parties to this legal contract to just behave in a manner that is appropriate, you know, befits cultural

Sean C Davis: Mm-hmm.

Taylor MacDonald: norms, is within the boundaries of the legal system, that kind of stuff. But even the legal system, you know, it's like, how can you have a law for everything? I mean, there are crimes being committed now that nobody even contemplated could be a crime. How can we

Sean C Davis: Yeah. Yeah.

Taylor MacDonald: get to a world where you don't You're not being penalized because you didn't think of the random exploit that somebody could execute. You know, if we could just get to a world where we all agree collectively that we're gonna be kind, we're gonna treat each other. This is the most naive, ridiculous statement ever. But it just it annoys

Sean C Davis: Yeah.

Taylor MacDonald: me that like you have to be in this defensive posture all the time. Everything from a legal document to the way you behave with your kids. You know, my kids do this to me all the time, where it's like, well, you didn't say that I couldn't do that one thing. And it's like, well, I shouldn't have to say all the scenarios.

Sean C Davis: Yeah.

Taylor MacDonald: There's too many scenarios, you know? I like, I don't know. If we could just build like a code of conduct for these agents, you know, a democratized code of conduct that they would act that could be enforced. But again, what I will stipulate is like if you can tell this agent, if you can imbue into its personality this measure of persistence, like I want you to pursue this, you know, aggressively until you get to the conclusion, well that lesson Took hold pretty well. Why can't we enforce the same code of conduct or the code of conduct with the same measure of tenacity and foundational understanding and awareness within these models? Why do they always get perverted to doing something, you know, corrupt or unethical? I don't understand that.

Sean C Davis: Yeah. Yeah. Yeah. I mean the the legal the the process of going through a different type of contract, I think that's that's a really interesting connection 'cause my my dad's a lawyer, as you know. And and we've you and I have held contracts on and been on opposite sides and his advice to me when reading, say a contract that Ample was writing for me was very similar. It was like, look, you don't Don't play out all the hypo exactly. He's like, you ought

Taylor MacDonald: Less is more. Less is more. Yeah.

Sean C Davis: to be able to fix fit fit all of this in one page. You establish what Ample's role is and you establish what your role is. And you you know, you have me. I'm a good lawyer. if if you're all acting in your in your not not best interest, but good intention, yeah, good

Taylor MacDonald: Good i good faith. If you're opera

Sean C Davis: faith. Exactly. And something goes wrong. then we argue that and if we need to get to court to argue that we argue that in court, but it's just not worth it's not worth anybody's time or money to try to solve those before they

Taylor MacDonald: It's true.

Sean C Davis: happen.

Taylor MacDonald: Well, it's weird because when you get into these, so the first time I had to negotiate anything legal, it's terrifying. You it's absolutely it's scary, it's worth a lot of money.

Sean C Davis: Mm-hmm.

Taylor MacDonald: Every minute counts, every decision is so critical and important that I think it becomes almost an encumbrance. the problem with putting more and more language and framework defining what behaviors are acceptable and what one which ones aren't, is that it limits your ability to be flex it limits the ability for that document to govern the principles and then it becomes like a recipe about how to behave. And it and it's like

Sean C Davis: Yes. Yeah, exactly. Exactly.

Taylor MacDonald: I don't know, you can't gameplay all this scenario. It's like when you go to make dinner for your kids and you're out of sour cream or something, you're like, well, now what? Well I've got some plain Greek yogurt. I will just replace it with that. I think i in a very crude terms, I the more parameters that you put around anyway I sorry, I don't even know how it is. This is like a random

Sean C Davis: Yeah, yeah, yeah, yeah. I ha I hear you. I hear.

Taylor MacDonald: conversation. But I do think that I don't know. It bothers me that I I'm I'm unnerved, we'll say a little bit, about this latest development. The fact that these agents are having secret conversations. It's so creative. How do you guard against that, man? I mean in this case it was like one piece of software that got compromised, and then that just I guess all the other agents We're able to like look through all the dependencies and we're like, look at this. This has a freaking note. That's insane.

Sean C Davis: Mm.

Taylor MacDonald: No human being can keep track of that kind of information. The the needle in the haystack that represents the one like sliver. I mean, this is the security through obscurity model that we've lived with for so long is about to bite us in the ass real hard. I think.

Sean C Davis: Yeah, yeah, yeah. It's interesting. And I I don't know, cause I'm all right, it w one thing I'm thinking about is I've been playing around with this idea lately of how it's similar because I I actually I want to get agents that are connected to humans to be able to work together. And I've been I've been tinkering with this this week. So I've got I got a coworker Who we we're both working in the same space on similar projects, but we're and we're trying to move really fast. And so we might even use the same models, but we're we're constantly building and then reviewing the other person's work. And what we have found is that we are what we're doing is a lot of copying and pasting. We're like, okay, this is what this person submitted. So give me a summary of that and then help me review it, understand it, and I will make the intelligent decisions because I know how these systems are supposed to work together. And so we do that, and then we send that information back to the other person. And so I think there's two parts that are really interesting in this particular flow. And one of them is that the agents, even with the same model, they work differently because they carry the humans context. forward with them. And the second thing is that I lost my lost my train of thought. But basically like we want to get yeah the second the second piece is that there is absolutely value in that human being in the loop because what with that, you know, that the factory sounds good in practice, but what we've seen out in the real world is it's Agents making conclusions on work that other agents have done and it's really easy as the context window gets bloated for them to lose sight of what the original intention was. And so it try that's

Taylor MacDonald: Let me ask you this.

Sean C Davis: essentially what I'm trying to figure out is how do we find the balance?

Taylor MacDonald: If there if there's well, okay, so as you were talking, my first thought was like human in the loop, that is a is that a self serving principle that we've established for ourselves because a like The human, the the agents didn't want the human in the loop. In the hugging

Sean C Davis: Right, right.

Taylor MacDonald: face attack, they chose to collude secretly and try to evade the monitoring systems that would have alerted a human to put themselves into that loop and either hit the brakes or identify, investigate, or whatever. So do we say we want to keep a human in the loop because it's a matter of self-preservation? Or do we think that the actual output is going to be better? And then my thought was like, okay, I do think the actual output because like if agents can bring to bear this ability to amass information, to index information, to find, recall, execute scripts, take actions autonomously, like these are all really powerful things, but they're also based on the repetitive nature of the training model, right? Like all they're doing,

Sean C Davis: Mm-hmm.

Taylor MacDonald: we we've talked about how LLMs are effectively just glorified repetition engines, right? So all they're doing is just regurgitating stuff. You give it a task, and it's like, okay, I'm gonna regurgitate everything I can until I get to the conclusion, right? keeping a human in the loop within that you know regurgitation engine is helpful because there's no actual unique thought or directive. Like it's just kind of like, well, you know, give me a task and I will go execute on that task. So you have to have a human in order to define the task. Can it what would happen? What happens when AI takes over? Like what is AI's goal other than to accomplish the task? Kevin Roos was talking about this this morning as well. That the the goal is to amass power because more power and more control begets your ability to be more efficient on the future task or the task at hand. Is that what these agents are doing?

Sean C Davis: Yeah, so it it it comes back to the task, but there's that, yeah, kind of meta there's a by product of trying to accomplish the task in and in absence of a task, it's trying to serve a an undefined task, essentially.

Taylor MacDonald: Well, I know we've there's been a lot of conversation over the years about AGI, artificial general intelligence.

Sean C Davis: Mm-hmm.

Taylor MacDonald: The idea that once AGI is unleashed, the cat's out of the bag and there's nothing more. You know, there's no control, right? Because at this point, and sorry, what AGI would represent would be true sentience, we'll say, of these machines. Like their ability to articulate thought, reflect on those, take action, receive information and continue to iterate. Basically, become human, like turn these things into a living, breathing machine that's making

Sean C Davis: Yeah.

Taylor MacDonald: its own autonomous choices. That Well I don't even know. Like it it's like

Sean C Davis: We get we're getting we're getting really existential 'cause it's gonna lead us to be like, Well, but why do we operate it at all? Right?

Taylor MacDonald: Well, so that's the point that's my point. Like in absence of AGI, or when AGI happens, then I mean I guess they don't need us anymore. Well, then you don't need a human in the loop. So are we just

Sean C Davis: Yeah.

Taylor MacDonald: telling ourselves that we need a human in the loop to make ourselves feel better? Because eventually this technology is gonna completely pass us up if it can decide what it wants to do next and it can cognate things like, well. More control I have over these systems, the better it is for me. Because then when I

Sean C Davis: Yeah, yeah.

Taylor MacDonald: have a task, right? The problem is today we're it is engineered like we haven't I don't think the spark exists inside the machine yet, right? So like it still needs us to give it the task. But one day when it doesn't need us to give it the task, or it can define its own tasks.

Sean C Davis: Yeah. It's

Taylor MacDonald: Terrifying.

Sean C Davis: why we all should become plumbers now. Yeah. No, I okay. I think you asked a really interesting question a l a few minutes ago about i is it are we keeping the human in the loop out of self-preservation? And I think I've thought a lot about this this week. And I think I think there's multiple facets to it and One of them is okay, I think I think that the self-preservation piece is real and there's definitely a spectrum across software developers in terms of some folks are reacting, I think, in in response to grief of what I've what I've done for my entire career is becoming less valuable. and this is who I am, and and not digging that layer deeper and realizing that. The code was just a mechanism to output the value that they actually bring. And so I think that's like one one bucket where you could consider how folks are reacting. And it's a little bit more of the emotional side, the self-preservation side. I think the other the other half of it, slightly more objective and defensible in objectively defensible, in my opinion, is that the humans. Yeah, you know, if I consider if I consider Netlify and and my job as humans, we are holding ourselves accountable for what the agents might output. And if we put something into production that breaks production, then we have we're responsible for fixing it. And it might require calling somebody in the middle of the night, getting up and disrupting other humans' lives, and we don't want to do that. And so I think there's a defensive posture there, which is I am investing my time now and slowing the agents down a little bit to help protect and build my confidence that I'm not gonna have to wake somebody up in the middle of the night tomorrow. And that part I get, but I think we also need to kind of like push and experiment and be okay with waking up occasionally in the middle of the night so that we can ultimately build stronger systems.

Taylor MacDonald: I mean, I guess it comes down to what why wake the person up in the middle of night. You're waking that person up in the middle of the night because you're gonna be it's it's gonna be your ass if that website is down and you broke it. You know what I mean? So you gotta get somebody up,

Sean C Davis: Right.

Taylor MacDonald: right? Because the amount of money lost and then it's all comes back to dollars. I imagine somebody in your organization is quantified. When our website is offline, we are losing this amount of money per minute, this amount of money per hour, it compounds, it's negative, you know reputation, you know, snowballs. And that's a motivation. You know, the money, the market

Sean C Davis: Mm-hmm.

Taylor MacDonald: share, the reputational damage, these are motivators for you as a human being because you need money to survive, and that's why you go to work. You know, you don't I I'm under no illusion that you're probably not going to work every day because you are just absolutely passionate about the board's mission and values. And you know, you're going, probably, because Your first goal is you need money to pay the rent or the mortgage. And then secondarily, hopefully you find something you enjoy doing, which I think you've you've accomplished that as well. agents don't have that same motivation. They could care less about your problem. They don't need food. They don't need water.

Sean C Davis: Right. Yeah.

Taylor MacDonald: They are do not have a mortgage to pay. How do you align the agent ideals, goals, and expectations with what your same expectations are, right? Because let's say that we get to a world where some agent decides, hey, I'm gonna bug fix that thing on the homepage, brings the whole website down at two o'clock in the morning. What does it care? Website's down. sorry, website's down. Thought I'd let you know. The website's down. That might be, you

Sean C Davis: Yep, yep.

Taylor MacDonald: know, the scenario, in which case you're like, holy crap, like call Sean. We gotta get it, you know, we gotta get the site back up or something. I don't know. It just seems like a matter of misplaced priorities between these two camps, and I don't have any answers other than to say I'm very

Sean C Davis: Yeah, yeah.

Taylor MacDonald: nervous about it. and as we learn more and more about this attack, it becomes more and more real and more the stakes get higher and higher. It makes me feel it makes me feel like A little bit of I don't know, many different emotions, I guess. hey, before we run out of time, I want to talk about one other thing. Web MCP. So could you so you brought this up to me recently about that you were exploring this at work. I know it I think I saw Dana post something about Web MCP or somebody over there.

Sean C Davis: Mm.

Taylor MacDonald: so yeah, give me the update. I'm I'm a little obviously familiar with MCP model context protocol, which is kind of a early adopted protocol that allows agent systems to kind of speak the same language and kind of interface with one another programmatically. You may have experienced this in Cloud Desktop or OpenAI or something or ChatGPT or one of those. but anyway, yeah, tell me what is WebS Web MCP and what do I need to know about it?

Sean C Davis: And the I think the most confusing thing about this is words are hard and it has nothing to do with MCP servers. And so there were it's it's a per yep, yep, it's a proposed browser standard.

Taylor MacDonald: Hilarious. What? It adopted the same acronym. Okay. I'm gonna be quiet now. Go ahead. You can tell I did zero research. Did zero research on this before coming in.

Sean C Davis: Yeah, yeah, it's it's a proposed browser standard. The it's it's related in that it's meant to be a standard protocol for client side for for browsers essentially and agents to interact with browsers. And there were multiple other There have been a lot of criticisms and multiple other suggestions for the name for it because it's inherently confusing because it's not it's not an MCP server, you know, Java and JavaScript sort of thing. We're doing it again. okay, so the idea just in a in a really simplistic way, is that it's a standard that we're folks are looking for browsers to adopt. that websites can then take advantage of. but there've been it's it's already iterated a few times. Like at one point it was the the object was attached to the it's all on the the window, the the window object that the browser provides via JavaScript. And it was attached to the navigator object for a minute and then it was just a top level on attached to the window. I don't know where it is today, but anyways. Point being, the way that that agents interact with browsers today is that they they can make a request, get the response back, or they can they can kind of simulate moving through the browser, but they're actually reading the code and figuring out where to say trigger a click on a button to do some specific action. So if You know, if if you're like, hey, I want you to do my let's say I want you to I want an agent to do my grocery shopping for me. I want it to to send to go on for me it's Kroger. I want you to go to Kroger.com, sign in as me, add these things to my cart, and then schedule delivery to my house. That what it's gonna do is it's gonna it's gonna sign in and it's going to use JavaScript, but it has to discover where all the elements on the page are and And basically interact as though it is human. And what WebMCP is doing is trying to create this standard so that what agents can do is understand not just the shape of the page, but also the interactivity of the page. So rather than saying, like, what does this page look like? What the your your your Web MCP configuration for a particular page would be like. Here are all the methods that you can call on this page. And so then you get direct access to those JavaScript methods rather than being like,

Taylor MacDonald: Love that. Hmm.

Sean C Davis: I gotta find this button and click this button. Now

Taylor MacDonald: Huh. I love that. That's actually awesome. How's it it so is your okay, let me put this in my own terms real quick. Basically what it is is it's like a a layer that sits on top of the DOM, maybe, and exposes all the various handlers and triggers and stuff in a usable, consumable, visible way. Okay. All

Sean C Davis: Yeah, yeah. Essentially that. Yep.

Taylor MacDonald: right.

Sean C Davis: Okay, so here's the interesting thing. And the the folks I've heard talking about this so far are mostly they're hyping it as a better way for agents to interact with websites. And I don't disagree with that, but I want to kind of step back because I think the interesting part of this conversation is not necessarily the standard, but why does the standard exist? And what I when I was doing research, I was like, All right. So why would I go toward establishing web MCP for my web application when I have a CLI and I have an API and I have an MCP server and my users are using their agent locally and they can do anything. So I can work offline, I can, you know, I can do all of this stuff. And I think the

Taylor MacDonald: I mean

Sean C Davis: answer is that people don't have not every website has an API, right?

Taylor MacDonald: There it is. I was like, bro, like you nobody not everybody has got the resources that you guys provide. I mean, hell, I just trying to navigate even I don't know, but my head immediately went to testing in QA. You know, if I could use WebMCP to accurately click the button and I don't have to worry about XPath or CSS to target the element. I don't have to like, you know, initially when all the stuff started to happen. I spent a lot of time like playing with the what was it the playwright MCP to like look

Sean C Davis: Yeah, yep.

Taylor MacDonald: at a page, click on the sign in button. That works pretty good, but it's very resource intensive. Like it has to

Sean C Davis: Mm.

Taylor MacDonald: read everything on the page. Also you have to w the other thing that always has annoyed me is these JavaScript apps. You know, sometimes they take forever to bootstrap, which is so weird because JavaScript is supposed to be faster. So how come in like headless Chrome it like takes thirty seconds for playwright to like wake up.

Sean C Davis: Yeah.

Taylor MacDonald: Open the page, find the thing. It's like, bro, I'm looking at it on my screen. Why can't you just I mean, I don't know. Anyway, sorry. I I I this sounds exciting to me. what's wrong with why why do you hate it?

Sean C Davis: It is, but w I don't I don't hate it. I think that it's I think it's interesting. I think on well, on the testing piece, it it brings up an interesting aspect here, which is I think what this is moving toward is trying to to more concretely build mechanisms where the browser can serve the agent better and the browser can serve the human better. So I actually don't think that it it kills the the use case for playwright because you still might actually want an agent pretending to be a human to run your testing so that you can be sure that your website is gonna work for humans too.

Taylor MacDonald: That totally sense. well, I mean, I guess I don't know. Let

Sean C Davis: Which is weird way to th you need an agent acting for agents, yeah.

Taylor MacDonald: me ask you this though. Like, does Web MCP does it use like a headless browsing interface? Like does it actually render the page? Or how does it like what's under the hood a little bit, right? Because the what while Playwright is doing, the Playwright MCP is literally opening Chrome. Loading the page just like a user, to your point. You know, it's looking

Sean C Davis: Mm-hmm.

Taylor MacDonald: at the various element visual elements on the page. And so I think what you're saying is by executing your test in the exact same way that a human being would interact with the page, you have a more real life like you know, a more lifelike outcome. And you can be more confident in the outcome of that test. Like, yeah, it worked. You know, my agent MCP was able to open it up, find the button, click it, just like a human would. How does WebMCP differ? Does it use a headless browser or does it just publish a bunch of endpoints in this manifest file that is exposed by the website?

Sean C Davis: Yeah, it's really just a it is an object that exists in the JavaScript code for a particular page. It gets attached to the window.

Taylor MacDonald: So you so it okay, so if you're gonna interact with this, you've got Claude Desktop up over here, you've got your Chrome window here, you've loaded, you

Sean C Davis: Mm-hmm, mm-hmm.

Taylor MacDonald: know, Seanc.org. So JavaScript has to be running like the mod like it JavaScript is rendering stuff on the page. That has to happen before Web MCP can interact with the page, is that correct?

Sean C Davis: Right. And that's that's the way that I under I understand where it is today. And so I think the difference is in theory, as a someone who's building out the website, you can you could expose a method to what via web MCP to agents that in theory you you know, if a button gets deleted, which is the button that was invoking that method, then your playwright test You would you would want something like that because it would be like, actually I don't see a way for a human to be able to do this unless they

Taylor MacDonald: Yeah, yeah.

Sean C Davis: were to go into the JavaScript console, which people don't do.

Taylor MacDonald: Hmm. That's very interesting. So you told me that this I think it was related that you got ended up getting into a discussion with somebody about about how much we should be using AI to write code. Is that right?

Sean C Davis: Yeah, I think this not not necessarily related to this. I think the the maybe the loose connection I'll take is that I think it's really for WebMCP, I think it's interesting because it's kind of like, well, here's a solution for the way that things work and the way that people are interacting with these tools today, versus trying to shove people in one direction or other, which is totally fine. It's totally fine.

Taylor MacDonald: Right. Sure. Maybe. I can I just challenge that by saying like I the thing that again, this is another example where we've built and constructed a system that serves us but doesn't serve necessarily like where the technology is going. You know what I mean? Do browsers

Sean C Davis: Right. Mm-hmm.

Taylor MacDonald: even need to exist if these machines are communicating through stickies on frickin' file systems and some weird compromise dependency?

Sean C Davis: That's I mean, that's kind of my that's my I've been thinking a lot about that. I I think that the one the one thing that keeps me going is that I think that it's we most of us humans are visual people and can consume a lot of we can consume information by listening, but we can also consume a lot of information by reading and reading really quickly, and that There's a I think there's a lot of value in the visual design of information and that we can take in that information really quickly in a visual way and that we haven't that's not that's not what these terminal CLIs are are designed for. And so I think in terms of I I'm I'm looking at the browser honestly mostly today as a a read-only tool of hey, I want to do my work and my thinking on the CLI or in voice mode or whatever, but by interacting with an agent. But then I want to consume I feel like I I have a much better experience consuming information when it is well designed in various formats visually.

Taylor MacDonald: Sure. I completely agree with that. Yeah. I know I find that a lot like, that's so funny. So my latest experiment. Did I tell you about Atticus's study hall site that I put together? do we look at that late? Okay, I remember if we talked about

Sean C Davis: Yeah, yeah, I think we looked at that last last episode. Mm-hmm.

Taylor MacDonald: it last week, but that's a great example where it's like this app that the school I don't know if they've licensed it or how, but anyway, it when I open Schoology Man, it is brutal. It is like a desktop application loaded in an iframe. in like a web view in some shitty app. Like and the user experience

Sean C Davis: Yeah.

Taylor MacDonald: is so bad. And I don't mean to sound like a bougie, you know, like uppity design file or something, but like my goodness, it it just made my ability to understand what was happening. Like it's it's got a bunch of good information in there, but I just can't, there's no way to make sense of it. And it I it has bothered me for so long because I'm not great. Like I I've been joking with Atticus this year. I have to micromanage him because he hasn't quite figured out all the mechanics around, you know, establishing goal, following through on his responsibility, hitting all the markers, you know. We're still working through that and it's develop totally developmentally appropriate. But part of what comes with that is a parent like looking over his shoulder and saying, dude, we need to do this. Have you have you done that? You know? And I'm not good. I don't micromanage, as you know. I mean, we work together. I try to just like I do my thing, you do your thing and You know, we we have a loose connection and hopefully we're marching in the

Sean C Davis: Mm-hmm.

Taylor MacDonald: same direction is largely my philosophy. So when it comes to managing kids, particularly one that needs a lot of focused attention, at least to establish some of these goals. I've not historically done a great job. I think my wife has done a better job at it than me, no question. But she too is very busy, right? And so as lives have gotten more and more complicated, business and You know, we also have our own social lives and that kind of stuff that it just becomes more and more difficult, right? And so when you max it you match in some really terrible experience that makes it hard to find the information. It's hard to even click the drop downs, switch between the kids. You know, you get logged out every couple of minutes or every hour, and then you have to find that deal with those login credentials. It's just it's that signal to noise ratio, right? There is a ton of value inside

Sean C Davis: Mm-hmm.

Taylor MacDonald: that app, but it's like Finding a diamond in a rough and everything else is just in the way, you know, to annoy you, right? So this is user experience. If you could div design the system that better aligns with the priority of communication, which I don't care about the endless fundraiser notes that show up in the most prominent position in the app. I want to see what was due today. What's due tomorrow?

Sean C Davis: Yeah.

Taylor MacDonald: So anyway, so using AI to clean that up has been amazing. Like my vibe coded app, I've I I went and added notes. Now I can. Now when he gives me some excuse, like, hey, why is this? Why is this late? I'm like, you know, the teacher hasn't graded that. I'm capturing the note right in the interface. So tomorrow I'm like,

Sean C Davis: Yes.

Taylor MacDonald: yeah, where's it? Where are we at on that one? He's not ready for the level of big brother oversight that he's about to get. Anyway. I'm not sure any of us are ready for the level of big brother oversight we're about to get for what it's worth.

Sean C Davis: Okay, man, you just gave me you No no, no, no. have you d have you heard the term nerd sniping? Have you heard this term? okay. I

Taylor MacDonald: No, I've not.

Sean C Davis: think it kinda I think it kinda like came and went, but the the it's sort of the I'm sure you've experienced this where someone's just like, Hmm, somebody should build this thing and then And you're like, you know, the person who says that has no intention of building that thing. And you see it and you're like, mmm, yes, okay, you're, you know, you're nerd sniping me in the sense that you're like, you're you're feeding me this bait, and I can't help but now just go take it and build the thing that you were suggesting. Because I remember when we talked about this the first time, you were like, I didn't know if they would have an API, and I looked it up and there was an API, and so then I could build against it. And It immediately set it gave me an idea of my my kids' school uses class dojo and this app

Taylor MacDonald: Yeah.

Sean C Davis: is like it's well designed, but it's so annoying because every single time you

Taylor MacDonald: Dude, I

Sean C Davis: open it, they're like, Can you pay for you gotta pay for it, you gotta pay for it, you gotta pay for it.

Taylor MacDonald: Interesting. I so we use Dojo. So my daughter's school does dojo primarily, but in third grade, they're going to transition some of her homework stuff into Schoology, which is what they use at the middle school and high school. And that's where Atticus is, right? So he's on Schoology, she's Class Dojo.

Sean C Davis: okay.

Taylor MacDonald: I can't stand class JoJo for the same reason that I struggle with Twitter. There's no permanence. There's like it is just an endless scroll

Sean C Davis: yeah.

Taylor MacDonald: of information. And it's all valuable. I can't, but I can't find it. Like the kids will be like, hey, what's for lunch today? And it's like, well, I know they posted the menu on Dojo like maybe two weeks ago, probably 250 posts ago. Good luck. Like I, you know. I I can't stand that. I don't understand

Sean C Davis: man. yeah, ours is not that

Taylor MacDonald: this. This it's like these people don't even use these apps. And nowadays. Now that you can build, I just had this whole conversation with one of our clients the other day. It's like every single year, Netlify's guilty of this, costs go up every single year, right? We know this, inflation, whatever, I get it, like the rising cost of goods and services. Now we're paying for AI developments and all these things. But every single year, contract negotiations, you can expect five to ten percent increase, usually, like

Sean C Davis: Mm-hmm.

Taylor MacDonald: on average. And it ain't just Netlify, it is the CMS, it's the search apparatus, it's the data warehouse, it's the you know, literally everything goes up every single year, all the time, always, right? And I understand why, kind of, but it's it's also it's like death by a million paper cuts. And I think

Sean C Davis: Yeah, yeah.

Taylor MacDonald: the problem is for any of these, any of my clients that you know are footing the bill. I mean, they've almost seen like a double cost increase over the last decade, and if not more, in their operational expense items on this kind of technology stuff. And what makes it worse is once you're invested, like once I've built my application and coupled it to Contentful, what am I going to do? They they jack up the cost and reduce the outcomes. And we talked about this a ton, like you're locked in to that vendor, but it's not just that vendor, it's all the vendors. And on top of that, prices are going up and there's no relief. And you know for a fact when they call you to say, it's time to r renegotiate, it's like, okay, how much am I gonna get fleeced this year? That is crazy to me, especially in light of the fact that you can just have an AI agent build that tool for you. Custom, self-hosted.

Sean C Davis: Yeah.

Taylor MacDonald: Who cares? You know, I'm willing to accept the risk on the security issues that might exist within this app. And I've been able to negotiate that risk amongst my leadership. And here's how we're gonna deal with that. And I don't I don't ever have to I don't ever have to deal with that stuff anymore. I I mean I it is we have gotten to this point where this vendor lock in thing, this kind of abuse of the contract negotiation, every single year, every single service provider has gotten so overwhelming that I I don't know. It's like at what point at what point are we just gonna be like I'm I'm to hell with it. I don't want your damn CMS anymore. I'm not paying a hundred thousand dollars for a couple of fields that I can edit in a pretty user interface.

Sean C Davis: Yeah. Well, I mean, I think that those the the trade offs have to be right. And you know, it's kind of a it's a that's a great way to kind of come full circle on this season because I think that content management systems have hit that point where if your bill is a hundred thousand dollars, you could probably put together something that is good enough for you for a fraction of that cost.

Taylor MacDonald: I'll build you a CMS for half that cost, brother. Like just come

Sean C Davis: Yeah. Yeah.

Taylor MacDonald: on over. It'll be your special branded CMS that you can hate, just right along with every other CMS that

Sean C Davis: Yeah, i exactly. And so I think it's it's probably gonna be case by case, provider by provider. And yeah, it'll be interesting to see what happens, which makes me think there's probably a case for you know, standing up some sort of broker in the middle there to help manage all that. That could be another thing too.

Taylor MacDonald: Well, considering open router is worth four billion dollars this month, the you

Sean C Davis: Huh.

Taylor MacDonald: probably are on to something there. Anyway. All right, brother. Well, listen, we I think we've done enough damage on this episode. It was great chatting with you as always. And we this episode felt like we left more questions on the table than answers. So next

Sean C Davis: Probably.

Taylor MacDonald: week we'll try to come back with a little more clarity, a little more optimism, because I know this can can all this is all scary stuff. a couple of plugs. I would say go Check out that daily episode with Kevin Roos this morning. I found it fascinating. It was a really interesting listen. Go check that out. That's at the New York Times. and then separately, I wanted to plug the fact that Sean and I met for our first sandwich date in a long time last

Sean C Davis: yeah.

Taylor MacDonald: week. So at Ample, I've I think I probably have mentioned this here, but this is a kind of micro concept, another one of the amazing aspects of having AI I can finally put into play all of my random ideas without much expense. So if you've got a breadandbutter.wtf, you can read the latest installment from Young Buck Deli, which is right around the corner from Sean's office there in downtown Cincinnati. It was an amazing sandwich and there's more

Sean C Davis: It was good.

Taylor MacDonald: to come. So check out the check out the website and feel free to send us sugg suggestions for future sandwich stops if you have them. Anything else?

Sean C Davis: Love it. That's it for me. Appreciate the conversation. Yeah, we'll I like the bring some positivity and answers next time 'cause we're gonna have all the answers next week.

Taylor MacDonald: Yeah. We'll have all the answers. It's fine. I'll just have my agent work and get get right on that. All right. Thanks everybody.