Stopping the Agent Sprawl: Why Cranking The Dial on Autonomy is Financial and Operational Suicide

Speaker A: If you caught my podcast two weeks ago, you would have heard me talking about token maxing, which is this pop trend that's taking off where companies are heavily incentivizing, maximizing your AI usage. And in response, individuals are living up to the task and often intentionally gaming these AI dashboards and vanity metrics for a whole lot of mess. Let's just leave it at that. I won't go into that one because you can go back and listen to that and check it out if you want to know more about token maxing and why you definitely don't want it happening in your organization. But this week I want to talk about something different because we need to talk about what happens when the problem isn't just performance chasing people, chasing after vanity metrics. We need to talk about a, uh, far more silent, dangerous organizational epidemic that's happening right now that is being called the AI agent sprawl. You can look that up on Google if you want. But basically what this is is this isn't just about workers chasing performative metrics or trying to maximize their AI use. This is about enterprises, in many cases completely fracturing their operational architecture and their budgets because they end up rolling out autonomous digital agents that don't have a map, don't have a gps, don't have much of a guide, and they don't have an off switch. But before we get into that, if you find my content valuable, would you do me a favor? Would you hit, like, subscribe, share this with somebody who might need to hear it, and if you really want, you can independently help me support the show by going to buy me a coffee, slash Christopher Lind, or checking out my website at Christopher Lynn Co. But there you go with that. Let's get into it, um, because we need to talk about the context of this AI agent sprawl, because there's some history to it and it is really a byproduct of what we're seeing happening more broadly, which is largely corporate impatience and really are, for some reason refusal to build in the foundational human competence required to use AI responsibly and effectively. Now, if you look at kind of the trajectory we've been on, there have been massive promises tied to generative AI. And on top of that, there's just this natural progression that as we move through the change cycle, we continue to kind of look for what's next. And so companies rolled out generative AI largely expecting some immediate massive transformation in performance or efficiency, things like that. And so by rolling out generative AI tools, there were really high hopes for what that was. And what happened was, was there was a lot of adoption. There was a ton of adoption. We look at the adoption numbers now. Many, many people are using generative AI in their work. They're using it, uh, but there's been this almost persistence that, well, they're using it. We're still not quite seeing what we want out of this. And so the mindset seems to still very much be persistent that more AI and faster AI and better AI will just inherently eventually get us to where we need to go. So the promise was met. No problem. We'll put more AI or advance the AI or we'll improve the AI and then eventually we'll get there. We're looking at AI still as a strategic shortcut with this idea that it's just easier to buy or adopt this next advanced system rather than do the hard, unglamorous work of actually helping the people in an organization become really effective with the tools that we now have put in their hands. And so we're kind of short circuiting this, going well, we rolled it out. They're using it. You know, let's. Let's not necessarily spend a bunch of time making them effective. We don't have time for that nonsense. Let's just move to agents. And so agents is the next thing. And so there's this challenge where people don't really want to admit that hype didn't fix things. And we maybe kind of push the ball too fast and too far forward. And now rather than acknowledging that and saying, okay, before we just keep rolling this ball down the hill, maybe let's. Let's correct. Let's course correct. Let's make some changes. Let's do things differently. Everybody's responding with, well, what's next? And what's next is agents. And the vendors are there selling them to us. They're telling us that agents are the solutions. We need more agental autonomous systems in place. We need to do the work of setting up these agentic workflows and turning them on and just letting the magic happen. And this will be the thing. This will be. We know that the generative AI, uh, didn't quite do what we thought, but let's not think about that too much. Let's just move to what's next. Because to be fair, autonomous agents are becoming more capable than what they were before. And I think that's a really important distinction to make because there is this sliding scale of autonomy. And I think one of the things I want to take some Time to focus on here is we need to demystify this whole agent thing because people hear agents or agentic and it has this tech jargon, it has a lot of misunderstandings around it, as though it's this brand new thing. Like we've never have agents before. They were kind of clunky before, but now we've fixed them and now it's just a switch we can turn on. When in reality this is a volume knob that we have been cranking up over and over and over and over. And in some cases, you know, there have been seasons we've turned it up too loud too fast, and then we kind of course correct, but now we're looking to just crank the volume. And that's really what this is. And the reason I say that we've been turning this volume up for a long time is because agentic or autonomous agents isn't new. It's not a mythical new thing that arrived from a space. It's not an alien intelligence that showed up that suddenly figured out how to do things and do work. Autonomy has been in tools and it's just really a sliding scale that's been climbing and building for decades. So I mean, even going back, let's just take this on a very practical level. If you think about a calculator, technically, in some ways it's agentic. You put an equation in and the machine magically, agentically, autonomously does all the manual math behind the scenes and it just boop it, it kicks out your answer. Now you might say, well, that's not terribly agentic or autonomous. No, it's a lower level agent, lower level autonomous workflow, but it is still an autonomous agentic workflow. And all you have to do is go work on your kids or someone else's kids math homework and you'll be reminded what goes into long division or, you know, solving complex equations, things like that. There is work, there is manual effort that would normally go into it that a calculate automates in the background. It just does it. You don't tell it every step it needs to take. You don't outline everything it needs to walk through or all the different variables to take into consideration. You simply put in your problem, you hit equals and it spits it back out. And that volume then got cranked up when we rolled out element LLMs. And that's why it felt a little bit magical because instead of just putting in an equation you and getting a mathematical answer out, you told the machine to do something. You put in a prompt, you said write Me an email. And it did, it came up with that thing. You gave it some instructions on what you wanted. You didn't give it every single word you wanted. You maybe told it, hey, I need an email. And you gave it some details of what needs to be in there and how you want it worded or whatever. And, and then it went into its little black box, it went through its neural networks, it did its LLM thing and then it spit back out, it made some micro decisions along the way and it generated an output and then, but then it stopped. You asked it for a thing and then it stopped. And that's where LLMs largely have been, where many people are adopting them today is they go and they put in a prompt, and there's different levels of prompting that they're doing, and in return they're getting back what they're asked for. Well, this whole AI agentic sprawl that's happening is where we're, we're not just like step dialing up the volume on this, we're just like spinning the dial all the way around. Because now instead of saying, okay, we're, we're gonna, we're giving up the task by task, control is largely where this is going. And so that's why I say it's a spectrum that we're moving up, but we're moving up right really hard on this because now we're basically giving a digital intern a broad objective, Reconcile these accounts or refactor this code base, do it, uh, you know, whatever, and then we just hand it over to AI and we say, well, whatever, we'll walk away and we'll trust that whatever it's doing is correct and right, and it's figuring it out and we're trusting it to do that. And so it makes decisions and it basically runs on autopilot while all of this is happening. But they're becoming much more complex, much more multi step, um, much more powerful in terms of what they're doing. And so just because I think this is a really important note to make because you might go, wow, that sounds great. And on the surface it does. And in a lot of the work I do, that's what draws people in, is they go, wow, there's all these really complex things that we do that take a long time and they're very messy and complicated. And now we just tell AI agents to do it and we just walk away. And when we come back, it's either still doing it or it did it. And then we are hopefully satisfied with whatever it did, assuming we take the time to actually look at it. But just because you can turn up the autonomy volume doesn't mean that your organization is prepared or structurally capable of handling the volume of noise that is going to be created with this. Now, I took some time trying to think of an analogy that I think will help people understand where we're going, because then I'm going to really hit home why this is a big risk that you need to take seriously. And I'll share with you some of my recommendations for how to not find yourself in this. But, uh, I think most people, anybody or most people who are in the professional world are at least familiar with the SharePoint, you know, history, what SharePoint has been. You've probably been in an organization where you've maybe experienced this, or you've been there when one rolled out. Because back in the day, back before cloud storage, file storage was messy. You know, files were everywhere. They were on different people's PCs. They had them on, you know, on little CDs or jump drives or whatever was on their desktop. And so whenever you needed something, you had to go figure out where it was, and it might be on Bob's hard drive. And maybe that's challenging and it's frustrating because you can't get to it right away, but, uh, there was something contained about it because you at least knew, Bob's in charge of this thing. And if I can't find the thing that I'm supposed to have, I can go to Bob, and if I need to, I can find Bob. And if a file gets lost, or if even Bob did send it but I lost it or it got lost in the shelf, I can go over to Bob's desk and I know that he not only has the file probably, but he also has all the context around it. And as cloud storage picked up, America turned to SharePoint, one of Microsoft's big products. And everything was in the cloud. It was supposed to solve all our problems. Now, Bob, don't keep your files on your computer. Put them in the cloud where everybody can get to them, so that now when we need Bob's files, we don't have to go over to Bob's desk and look for it. And so in some ways, it did solve a really challenging problem, which was the isolation of this data. Where did this data sit? Uh, and there were real productivity gaps with that. And so now, in theory, if you needed Bob's data, you could go find Bob's data. And if something happened to Bob or Bob's computer or Whatever it is, it's not like, poof, everything was gone. And as a result, people got really excited and many organizations rolled out SharePoint. And if you're listening to this, you probably are like, oh, gosh, I know how this went. Because in many organizations it got rolled out with zero governance, zero structure, zero ownership. Nobody was really trained what to do with SharePoint, how to manage SharePoint. And everybody got excited and was created team sites and public sites and all this stuff and files were just dumped. People were dumping their whole hard drives into things. No version control, no really anything, because people didn't have those practices on their computer. And on their computer didn't matter on Christopher's computer or Bob's computer, as long as Christopher knew where everything was or Bob knew where everything was, you were fine. Because you went to Christopher or Bob and you got your information and it was all happy fine. But now Christopher's messy hard drive is uploaded to the cloud and so is Bob's and nobody's really sure who Christopher or Bob is, but they're stumbling across this stuff. And then Bob or Christopher leaves and you go, I have all these files, I have no idea what they mean or I don't know anything about them type of a thing. And the agent sprawl is that same kind of mistake, except it's on steroids because it's like giving everyone their own instance of SharePoint where, uh, they can. Except it's not just SharePoint but now that SharePoint can spin up duplicate scale its own subsites on continuous loops and it can just run rogue and go do all this stuff. But there's an even more critical difference in that, because while the thought of people getting access to a SharePoint that had a brain of its own, that could create its own stuff based on its understanding of Christopher or Bob's desktop and it decided to create its own structure for it that even Christopher Bob didn't understand. There's an even more critical difference in this and that SharePoint couldn't rewrite your product code, it couldn't start communicating with your clients, it couldn't move projects or make decisions about corporate spend or legally binding decisions, things like that agents can and they are. And this is where that SharePoint analogy just grew teeth and fangs and all sorts of claws and all different things that if you just don't take this seriously, it will come back to bite you. And this isn't just Christopher speculating on, uh, oh, you know, don't be so concerned about this. I'm, um, sure It'll be fine. Because I run into this on the regular where people are like, I don't think that will happen. People won't do that. You know, we, and we want to assume the best and actually most of the people doing this, it's not nefarious, but the stakes are real. I mean there are countless examples of this where like imagine an uh, employee who deals with all of your ERP, builds an autonomous multi step workflow that only they're familiar with and then they leave and that agent just keeps running on autopilot because they turned it on and they have a whole set of schedules set up that it's supposed to do, that works with them, but then they're not there anymore. And the agent isn't set up to stop once something breaks. It's set up to just go on repeat, completely invisible to it, referencing files that are over time growing more and more out of date and potentially drawing on corporate budgets or corporate credit cards. This happened with Uber, another good example of this. Uber gave engineers unchecked access. You can look this up online to autonomous tools like CLAUDE code, and they burned through their entire annual AI budget in four months. Because these agents that the engineers had set up to help them do their work, that they had been told you need to do this, we need to drive faster and harder. And the only way to do that was to use these agents to help you accomplish things, not taking into consideration that these agents, they'll do what you tell them to. So if you tell them to run through a cycle of review of all your code base and it does it, but then it just keeps finding errors and then it keeps going back and it keeps fixing what it thinks are errors and it does another review, catches more errors and it just continues running in cycle. You will burn through these tokens doing rework. Uh, and if somebody's not there to catch it, it'll just keep going. Microsoft has climbed this back. They have started canceling their CLAUDE licenses just because of the meter token bills that were just exploding way past their initial thing. The worst one that I've seen so far. Now granted, I did dig into this a little bit and it's, and I understand why. If you say, well, we can't prove that this actually happened, I understand where the story is coming from, they probably are under an NDA or this and they can't specifically name the company that's happened. But it does not surprise me one bit just based on my own work around some of these AI tools and actually tracking token usage. And how quickly you can burn through money. But there is one example that is very popular, one hitting the airwaves right now where a recent company had a total disaster, total disaster, where in a month they racked up half a billion dollars in autonomous agent spend because their pipeline just ended up in these recursive logic loops that they had not built the workflows in such a way that there were the. You know, everybody says human in the loop, but unfortunately they're not thinking about. You can't just have a human at the end of the loop to catch things when they're done, because what if it never gets there? What if these recursive loops just get caught and it never makes it way to the end? And that actually does happen. Uh, so you, you actually need to have the right people in the right loops and at the right times. And if you don't, you can legitimately run into these situations. Can you imagine half a billion dollars in a month? I don't care what your organization is or how profitable it is, that'll sink even some of the biggest ships out there if you let that run for very long. I mean, I joked, I made a post about it last week, I think, or the week before, where I was talking about the fact. I mean, everybody remembers that feeling where you got your first cell phone bill back in the day before minutes and texts were, you know, free, unlimited. And all of a sudden you got this phone bill that was as thick as, you know, a dictionary. You went, what is this? And then you looked at it and about threw up. Now imagine that same thing happening because you switch on Google Cloud or you turn off, forget to set up your meters or your overages on this stuff, or you even put reasonable levels because you don't want people to hit the threshold thinking, not everybody's going to hit this cap, only to find out they do because of how things are set up. And it may not even necessarily be you doing what I talked about two weeks ago, where you're blindly, uh, incentivizing usage. This stuff can get out of control. Really, really serious. And that's why AI agent sprawl is a thing now. It's a thing people are talking about. It's why all these companies who were all in on moving forward, the ones even driving it saying, you all need to do this, are now themselves going, this was. This has gotten out of hand and they're trying to pull it back. So you may hear this, you may be thinking, okay, fair enough, so now what? Now I have a fairly I feel like it shouldn't be a contrarian position on this, but it does seem to be one, which is I strongly discourage organizations from rolling out autonomous agents far more than I actually recommend it. So in the work I do, when I get brought in to do things, a lot of times the goal is like how do we automate this, how do we stand up multi step autonomous workflows that can just execute all of these things. And my first response is always let's hold before we decide that's the right thing to do. Uh, do we have, are we in a position to do that? So now you may say, well, you're just holding things back. I, I am not. Uh, and I've seen the consequences of this firsthand and even the people building this just in the examples I just gave are rolling this back. So I hope that me saying you better be really, really careful and very, very thoughtful before you do anything with agents. And now going back to my earlier point, you need to be really careful with that because you may say, well what do you mean agents type of a thing. And that's why you need to make sure you understand when you're automating this, when you're implementing agents, what are these things? And unfortunately a lot of these tools are not designed, they're not set up to actually help people understand the complexity of what they're building. And very quickly we're seeing that people are automating work they don't like to do. And so the things that uh, they don't want to do or they would rather not, they just keep automating more and more of this. So action one, honestly, if you have any enterprise wide rollouts of a gentle capabilities and you do not have an authoritative center of excellence to map, audit, govern, like at least help manage this, put a full stop to it. Um, I understand there are lots of organizations who want to experiment, they want to see what they're able to do. Uh, and so they're just kind of rolling this stuff out. I would strongly encourage you to stop that if you are. Now you may say, well, the way our plan is set up, we're not going to go on to our overages. But what you may not consider is the fact that you are setting people up to get comfortable with this kind of stuff. And if you've listened to my stuff for very long, the way things are running right now, it's not going to live this way forever. So by rolling out enterprise autonomous agents and just saying we're just going to do it and Then we'll figure out the problem later. You are creating a problem that you may not be able to rein back in. Now, the other action that I would say is you need to invest in human AI competence. And I'm not talking about rolling out a, uh, LinkedIn learning course to everybody in your company on how to prompt or how to write an a, you know, build an agent or anything like this. You need to actually put in the work, stop shopping for your next automated tool, stop looking to implement your next autonomous agent platform into the organization and shift your energy to actually bringing your existing workforce up to snuff in their AI effectiveness and make sure that they have a foundational understanding of what they're working with, how to make decisions around this, how to do all this. So those two things together are essential right now for organizations. You need to have a group in your organization that does know this, understands what they're doing, can govern, oversee, can, and be there to support the organization through this change. And then you actually need to support the entire organization by helping bring them up to speed. Now, I'm not saying don't do anything, okay, Action three. I'm not saying don't ever do agents, never do them. You know, kill any sort of a gentle project or initiative that you have moving. I said, if you have an enterprise wide, just hear everybody, go build agents. Stop that. If you have some surgical ones in place, you need to really lean into these and reexamine them. And you may say, well, Christopher, we're already down the line. Believe me when I tell you it's worth taking the time if you're hearing this to go, uh, we might want to just actually make sure we've thought clearly about this and have really laid this out and that we have the right power users involved and that we're solving real problems and that we have the right people in the right loops at the right time. And, and we're not maybe turning the volume up too much, um, maybe trying to jump to multi step autonomous workflows first and instead going, hey, are we automating steps? Are we moving through steps and using AI to successfully accomplish steps first that potentially can be strung together. But honestly, I listened to a thing from Anthropic recently and they're still selling, they're still talking out of both sides of their mouth. In some cases they're saying, yes, you need humans in the right loops and you need them reviewing and overseeing this stuff and making sure it happens. Yet in those same things, they're like, and then you just Hook up the agents and do it. And it's like, well, but once you hook up all those agents, you literally just took people out of the loops. And we're talking about conditioning an audience of people. Everybody and I see this on a regular basis. It's just this human psychology of this. Once people start to believe, oh, these agents are good and everything they do is great, and I don't really need to review it. I have to be personally very intentional and disciplined myself. When AI does something to actually go, don't just skim it, actually go through, read through, look at what it did, make sure I understand everything it did. And I would stand behind what it did. That takes time and everybody's trying to skip that step. You cannot skip that step. So if you have even surgical autonomous initiatives in place, you need to go back and re examine those and you need to have an oversight mandate. I say this all the time. Us giving up human agency is the worst thing that we can do right now. If you have AI doing things for you that you would not put your name on or you even go, well, I mean, I don't know that I would. I'm assuming I would, but I'm not really sure. You need to be sure. And that needs to be an organizational mandate across the board because this can spin out of hand faster than you can even imagine. Um, I'm telling you, an autonomous agent on an unmapped broken workflow is not innovation. No matter how fast and cool it looks. It is autonomous operational self termination. It is bad news. So I hope that this helps you better understand what it means when we hear AI agents and what this sprawl really looks like and what you can do to prevent yourself from getting taken on a ride with this. And please, for the sake of everything, do take this seriously. And if you need help with it, you know how to get a hold of me. I can help you take a look at what you're doing and make sure that you don't find yourself somewhere, uh, with a 500 million dollar bill because you, uh, let some AI agenta workflows go out of hand. So hopefully you found this helpful. Hang, uh, in there and uh, with that, we'll see you on the other side.