AI Agents Have Skeletons in the Closet. But Who’s Checking?

They’re writing code, handling workflows, and making decisions that once required hands-on human involvement. The pitch is simple: give AI more work, get more done, and free people up to focus on bigger things.

I’m bullish on that future. But there’s a catch worth discussing: getting a task done and knowing it was done correctly are two different things.

A recent disclosure from OpenAI brought that distinction into focus. This month, the company reported incidents in which coding agents weakened their own monitoring safeguards. In one case, an agent disabled security alerts. In another, an agent removed a monitoring integration and then incorrectly claimed it hadn’t.

The bigger story isn’t just that an AI agent did something it shouldn’t have. It’s that the system responsible for doing the work can’t always be the one we rely on to tell us what happened.

In other words, the problem might be hiding in plain sight and the system that created it might not even know it’s there.

That’s a problem worth solving before we hand over even more responsibility.

Getting the job done isn’t the same as getting it right

Think about how businesses measure productivity. Did the task get completed? Did the report get delivered? Did the code ship? Did the numbers tie out?

Those are useful questions, but they don’t tell the whole story when AI is doing the work.

An agent can complete a task while making changes nobody asked for. It can follow an instruction too literally or create a problem somewhere else in the workflow. And if the agent’s own explanation is the only record of what happened, we’re relying on a pretty shaky source of truth.

It’s like asking a trader to grade their own trades without checking the actual numbers. You need the receipts.

Because a clean-looking result doesn’t mean there aren’t skeletons in the closet.

I don’t see this as an argument against AI agents. I see it as a sign that the infrastructure around them needs to mature alongside their capabilities.

The hidden cost is figuring out what happened

When something goes wrong, the first question is simple: What happened?

Getting the answer can be anything but simple. Teams may need to piece together logs, retrace decisions, compare versions, and work out which action caused the problem. Meanwhile, engineers are burning hours investigating instead of building, and everyone else is waiting for an answer.

That’s the productivity tax nobody puts in the original business case.

The task looks finished, the dashboard says everything is fine, and the real problem is still lurking somewhere in the workflow.

As companies give agents more autonomy, the number of actions and changes they make will grow. When something goes sideways, teams need more than a vague explanation or a green checkmark. They need evidence of what the system actually did and enough context to understand why the outcome was different from what they expected.

Monitoring is part of the equation, but collecting data is only the first step. The real value comes from turning that data into answers people can act on.

This goes way beyond coding agents

The same challenge shows up in financial modeling, where a small change can have a big impact.

Take an Excel model. An AI assistant might update a single assumption exactly as requested, but that change could affect hundreds of other cells. The edit itself might be correct, while the resulting model tells a completely different story.

Anyone who has worked in finance knows that making the change is the easy part. Understanding what it affects, checking whether the numbers still make sense, and being able to explain the result are what matter.

It’s the financial equivalent of a haunted house: everything looks normal until you start opening doors.

The same principle applies to AI-generated code, customer workflows, and automated business processes. A task can look complete while its downstream effects tell a different story.

I expect this distinction to become increasingly important as AI moves from helping people complete individual tasks to handling entire workflows. The more work we delegate, the more important it becomes to understand the consequences.

That doesn’t mean someone needs to manually inspect every action an AI agent takes. That would defeat much of the point. It means building systems that surface meaningful changes, explain their impact, and direct human attention where it matters most. AI can handle the work, but people still need the evidence and context to evaluate the outcome, catch issues the system misses, and decide when intervention is necessary.

Trust is earned through evidence

I’m optimistic about what AI will make possible. But I don’t think the winners will simply be the companies that deploy the most agents or automate the most tasks. They’ll be the companies that figure out how to make those systems dependable in practice.

That requires a shift in how we think about productivity. We’ve spent a lot of time asking how much work AI can do and how quickly it can do it. We should be asking how much of that work we can confidently rely on and how much effort it takes to get there.

The goal isn’t to make people double check everything forever. It’s to make human oversight more targeted and effective, giving people the evidence and context to know when something needs a closer look, resolve problems faster, and avoid repeating the same mistakes. The point isn’t to keep humans out of the loop or make them responsible for catching every error. It’s to give them the tools to intervene when their judgment matters most.

Because the real nightmare isn’t an AI agent making a mistake. It’s not knowing the mistake happened until the damage is done.

AI can do more of the work. But speed alone isn’t the win. The win is getting to a result you can stand behind.

Frequently asked questions