What 'AI agents' actually do — and what they still can't
Every company now sells an 'agent.' Underneath the marketing, most of them are doing one of three things. Here is how to tell which.
The word agent has been attached to so many products this year that it has almost stopped meaning anything. A calendar app has agents. A customer service widget has agents. Your email client, apparently, has several.
Strip away the branding and nearly all of them fall into three buckets. Knowing which bucket you are looking at tells you most of what you need to know about whether the thing will work.
1. The one that fills in a form
The simplest kind takes something you wrote in plain language and converts it into a structured action. You type “move my Thursday meeting to Friday morning,” and it maps that onto an API call your calendar already supported.
This works well, and it is genuinely useful. It is also the least impressive of the three, because the hard part — the calendar API — already existed. The model is a translator sitting on top of it.
2. The one that runs a fixed sequence
The second kind chains several steps together: read the document, pull out the numbers, put them in a spreadsheet, email the result. The steps are decided in advance by a developer. The model handles the messy parts in the middle.
These are reliable in proportion to how narrow they are. A tool that only processes invoices from one vendor will rarely fail. A tool that promises to handle “any document” will fail often, and usually without telling you.
3. The one that decides what to do next
The third kind is what most people picture: the system is given a goal, and it chooses its own steps. It might search, read a page, change its mind, try something else.
This is where the real capability is, and also where the real failure mode is. The system has no reliable sense of when it is wrong. It will pursue a bad plan with exactly the same confidence it brings to a good one. Left unsupervised over many steps, small errors compound.
The test that cuts through the marketing
When a product claims to have an agent, ask one question: what happens when it is wrong?
If the answer is “it shows you the draft and you approve it,” you are looking at a useful tool with a human check. If the answer is vague — if nobody can tell you how errors surface — then the product has been sold before it was finished.
What this means in practice
The technology is real and improving quickly. But the gap between a demo and a dependable tool is mostly not about the model. It is about the boring work around it: what the system is allowed to touch, what it must ask permission for, and how mistakes get caught.
Products that have done that work tend not to advertise the word agent very loudly. They advertise what the thing actually does.