Turtles All the Way Down

AI will not run your operations for you, because someone still has to run the AI.

Photo by Wolfgang Hasselmann on Unsplash

There is this old expression, turtles all the way down. It comes from the idea that the world sits on the back of a giant turtle, and that if you ask what holds up that turtle, the answer is a bigger turtle, and under that one is a bigger turtle again, and it goes on like that forever, with no last turtle at the bottom actually holding the whole thing up. People call that an infinite regress, a chain where every answer just raises the same question again.

A friend brought the phrase up recently, and it stuck with me, because it is exactly what the pitch for AI running operations sounds like. We are sold a story where AI agents will deploy our code, watch the systems, catch the incidents and even fix them, and us, humans, can finally get rid of the pager and get back to sleep. It is a good marketing story, but I have never quite bought it, because it makes the same move as the turtles. It assumes the stack can go all the way down without ever reaching a floor. But let me ask you this: who runs the AI?

To be clear, I am not talking about the yoga studio that asks an agent to build a website and watch its sales. Nothing changes for them, and nothing needed to, because they were renting that website long before any of this and never owned the stack underneath it. They can consume AI-run operations the way they already consume the cloud. I mean the layer below, the software companies whose actual product is the running system. That is where the promise really lands, that AI will finally let the operators go home, and that is the promise I want to challenge.

The models are software too, and there are really only two ways to get them into your stack. Neither one makes the problem or the pager disappear.

You can outsource the frontier model and access it behind an API. You do not run the thing yourself, which might seem like a good idea, until you notice that your critical systems now depend on a company you do not control and, often, a government that is not yours. That provider can change the model under you, deprecate the version you built on, price you out, or simply be instructed by its own government that you can no longer access it. We have already seen access to frontier models drawn along national lines. For a normal product that is a risk to manage but for your critical infrastructure, it is a very serious one, because you have now moved the bottom turtle offshore, into someone else’s jurisdiction and someone else’s politics. And there is a plainer dependency sitting under that one. A hosted model is only useful when you can reach it, so your operations now depend on the internet being up. The AI that is supposed to run your systems and pull you out of an incident lives at the far end of a network you do not own, and incidents have a way of taking the network with them. So the moment you most need that model is often the exact moment you cannot get to it.

Or you can own it and run the models yourself, which many companies seem to be starting to do for privacy, for cost, or precisely to avoid that dependency. So now you are the platform. Instead of managing access to a single frontier model, you now manage access to a dozen or more small task-specific ones, and you are suddenly in the business of GPU capacity planning, model versioning, inference that has to scale under spiky load, weights that are now a prime target to protect, and keeping all of it patched and observable. That is the exact operations you were trying to remove, handed back with a bigger bill. The agents were supposed to shrink the operational surface. They enlarged it, or they moved it to someone you now have to trust.

Either way you do not escape the problem. You either run the operations yourself, or you hand the dependency to someone who does and hope they stay reachable and on your side.

At this point, people tell you that AI will run the AI, and that is exactly where it becomes turtles all the way down. The AI operating your infrastructure is itself infrastructure. Point a second agent at it to keep it running, and you have not removed the operations, you have moved them up a level. Point a third at the second, and you have moved them again. It never reaches the ground.

It cannot reach the ground in AI, because that would be a circular dependency. If the thing that recovers your systems is running on the systems it is meant to recover, then the moment the base goes down, your recovery goes down with it. Anyone who has run systems knows this and has learned it the hard way. You do not build your recovery on the part most likely to be down, and an agent should not get a pass on that. So the bottom of the stack has to keep working when everything above it fails, and in practice that means a human.

And in practice it is worse than a point of principle, because of how concentrated all of this mess is. The vast majority of the models run on a handful of clouds, the same clouds your systems already sit on. So the AI you would lean on to operate and recover does not stand to the side as an independent thing, it shares its fate with everything else. When one of those providers has a bad day, and it will, it won’t just take your service down, it will take down the reasoning layer you handed your operations to, at the exact moment you need it most. And when the work is handed to an agent, your on-call engineers stop practising the very skills you would need in that moment, and skill atrophy is real. When us-east-1 last went down, it took a good slice of the internet with it. Back then that meant your users could not reach your app. The version we are building means the next time it happens, nobody, human or machine, is in any state to think. That is the part that worries me.

The obvious counterexample here is the internet. We are told that AI will become a commodity, just like the internet, and that nobody at a normal company thinks about who keeps DNS running or how the undersea cables get repaired. A small number of specialist companies run that internet base, everyone else builds on top, and it just works. So the argument goes, AI could be the same, a few companies holding the models and the hardware, and agents doing everything above it.

I hear that commodity argument a lot but it never sat right with me. These layers do something very different. The internet is transport. It carries your business from one place to another, it does not run it. DNS does not decide what your company should do next, and TCP has no view on your customers. You can forget those layers precisely because they only move things, and moving things reliably is pretty much a solved, boring problem. AI is not being put in the transport layer. It is being put in the business layer, or at least the decision layer, the part that chooses what to do, what to deploy, what to roll back, what an incident means and what to do about it.

And a decision layer cannot be trusted and forgotten the way transport can, because when transport fails you lose a packet and retry it. When the layer making your decisions fails, you do not lose a packet, you lose time. It has confidently made the wrong call, and your customers stay in the outage for as long as it takes you to work out that it did.

So AI does not take operations away. It adds a layer. Now you run your own systems, and the fleet of models, and the agent platform that orchestrates them, and the governance layer that is meant to keep all of it in check. That is more to deploy, more to secure, more to keep available, and a great deal more that can fail. So much for stepping back from the pager. The operator does not disappear, they move up a floor and get handed a bigger, stranger stack to keep alive.

You might say the humans are a stack too, the operator under a manager under a board. That is fair, and a friend of mine made that point, he called the human floor a different looking turtle. He is right, it is another turtle. But it is a different kind of turtle, and that is the whole point, because a person can be accountable for a decision and a model cannot. So the recursion does not stop because you run out of layers, it stops at the first turtle that can look at the result, take responsibility, and say stop. Every layer above that one can only pass the message along. Only a person can own it.

None of this is an argument against AI agents for operations. I am sure they will do real important work, there is plenty of it. Instead, it’s an argument against the part of the story where we are told the humans walk away, put down the pager, and go back to sleep. You can push the operator higher up the stack, you can give them better tools, you can automate the dull middle. What you cannot do is remove the bottom turtle, because everything above it depends on someone being there, on-call, when the whole thing comes down.

We have watched this pattern before. Kubernetes was going to make running systems simpler, more portable, easier to scale, and it does some of that. But it also turned out complex enough that operating Kubernetes became a specialty of its own, a role and often a whole team that did not exist a few years earlier. The promise was fewer operators. What it delivered was a new kind of operator. Agents make the same promise, one level higher, and it will land the same way, more to run, one more turtle, and a human at the bottom holding it all up.

//Adrian