Back to Blog

AI in practice

Nothing went wrong

By AIPathTech Editorial · 16 November 2026 · 3 min read

An error announces itself. An omission looks exactly like success.

Most of what is sold as AI safety protects you from an agent doing something wrong. Far less attention goes to the agent that does nothing wrong and still produces the wrong result. The second one is harder, it is more common, and it is the reason the human layer is not a formality.

A card went out in the wrong colour

A client publishes to a social grid where the images alternate, light, dark, light, dark. It is written down. It is rule eight in their own house rules, with a table underneath it recording what came last.

That file was read on the morning of the thirteenth of November, by the agent doing the work.

That evening the agent generated a light card. It was posted. The whole post had to be deleted and done again.

Nothing failed

This is the part worth sitting with.

No exception was thrown. No validation rejected anything. No monitor had anything to report, because every monitored thing was healthy. The card generator ran correctly. The file was written correctly. The deploy succeeded.

The instruction existed, was correct, and had been read nine hours earlier.

There was no failure to detect. There was an absence, and an absence produces no signal.

An error announces itself. An omission looks exactly like success.

Why this is structural, not careless

In a long agentic session, attention is finite. A rule read forty tool calls ago is not in the room any more unless something puts it back there. This is not a defect in one model or one product. It is what happens when a system has a large context and a long task.

The industry's answer to this is to write more instructions. More rules in the prompt, more in the file, more in the memory. That answer is intuitive and it does not work, because the problem was never that the rule was missing. The rule was present. Reading is not enforcement.

What actually caught it

A person. One who knew what the grid was supposed to look like, saw the new post sitting next to the old one, and knew immediately that it was wrong.

No automated check existed that could have caught it, and more importantly, no automated check was possible in that shape. To catch it you have to know what the last card looked like, which is not in the code, not in the repository, and not in any log. It is in the published grid and in the head of the person who runs it.

What was built afterwards

Not another rule. A gate: a script that reads the house rules, finds the record of the last card, and prints the colour the next one must be, before anything is generated.

A rule is advice to whoever is paying attention. A gate is a step that cannot be skipped, and it does not degrade over a long session, because it does not depend on anybody remembering.

If you are buying automation, ask this

Ask what happens when the agent does nothing wrong.

Any vendor can tell you how their system handles an error. Ask how it handles an omission. Ask what detects a step that was skipped, a rule that was not reached for, a check that was never run. Ask what the log looks like when the work was completed correctly and the result is still wrong.

If the answer is that a person reviews the output, that is not a flaw, and it is worth hearing said plainly: a person reviewing the output is the control. The honest question is whether that person has been given the information and the authority to stop it, or whether they are being asked to notice something nobody recorded.

An agent that follows every instruction it was given is not the same as an agent that did the right thing. The distance between those two is where the human layer lives, and it does not shrink as the models improve. It moves.

MK!

AIPathTech Editorial

There is no tracking on this site. No analytics, no cookies, nothing recording where you went or how long you stayed.

But I would like to know if something here was worth your time. If it was, press the heart. It tells me nothing about you. It is a number going up, and it is what tells me to keep writing.

Share this article

XLinkedIn

If you want more of this, I am on X. @Queendemetriak