Software, Tasks — and Permission to Act
Part 2 of 4 — Understanding the technology behind the headlines.
AI ExplainedYour email client has been checking for new messages for years. It connects to a server, fetches your mail, and lets you know when something arrives. You don’t have to tell it to repeat the job every few minutes.
Your backup software does something similar. So does a scheduled script that checks whether a website is still online.
Software acting on our behalf is nothing new. Yet add “AI” to the description, and suddenly it sounds as if we have invented a digital employee.
What actually changed?
From following rules to asking for help
Imagine you want your email client to put important messages at the top of your inbox.
You could give it a rule: anything from your biggest customer gets marked as important. That works, but it also flags their Christmas greeting while missing an urgent message from someone you have never met.
A model gives the software another way to handle the task. It can assess the content of a message, rather than just check the sender against a list.
The software retrieves the email and asks the model to judge its importance. It then uses that answer to organise your inbox.
This is a straightforward example of software using AI. The overall job still follows a fixed sequence, but one step now involves a model’s judgment.
What makes it an agent?
Now give the assistant a broader task:
“Get me ready for tomorrow’s meeting with this customer.”
There may be no single list of steps that fits every meeting. The assistant might search recent emails, find a proposal, read an attachment, and check whether the customer replied to your latest offer.
What it finds determines what it does next.
That is a useful way to understand today’s AI agents: software that uses a model to help choose and carry out the steps needed for a task. People use the term differently, but this ability to adapt the next step is an important distinction from a fixed routine.
The model might decide that an attachment is relevant and request it. The software opens it and returns the contents. With that new information, the model can decide whether to keep looking or prepare your briefing.
There is a conversation happening behind the conversation you see.
The model needs a way to act
A model can produce the words “I sent the email.” Those words alone do not send anything.
For that to happen, the application needs a connection to your mail service and access that allows it to send messages.
The model can choose an action and request it. The surrounding software performs the operation using the tools available to it.
So yes, a model can help direct what happens next. But its ability to affect your inbox, calendar, or files comes from the system it has been connected to.
“AI agent” is a convenient name for that whole arrangement. It is still software using AI, with a task to complete and access someone has granted.
Permission changes everything
Consider two assistants using the same model.
The first can read your email and prepare draft replies. You review each one and press Send.
The second can send replies without asking you.
They might write equally good messages. But the consequences of a mistake are very different.
The same applies to files. An assistant that can suggest which files look unnecessary has a different role from one that can delete them.
More permissions do not make an agent more intelligent. They give its decisions a wider reach.
And telling an assistant “please don’t delete anything” is different from giving it access that prevents deletion. The first depends on following an instruction. The second puts a limit on what the software can do.
What about several agents working together?
A task can also be divided between agents. One gathers information, another prepares a draft, and a third reviews it.
They may all use the same underlying model, with different instructions and access to different tools.
That can be useful, especially when the work breaks naturally into separate jobs. But agreement between three agents does not automatically make an answer correct. They may all be relying on the same mistaken source.
Dividing the work helps organise it. Checking the result still matters.
Start with the job
An agent is useful when it saves you from repeatedly deciding what needs to happen next. But plenty of tasks already work perfectly well with ordinary automation.
Checking for new mail every minute does not need a language model. Preparing a briefing from a messy collection of emails and attachments might benefit from one.
Before handing either task to software, ask what it needs to access — and what it is allowed to change.
That tells you much more than the word “agent” on the box.
Next: “AI did it” — what actually happened, and why the distinction matters.
Top comments (0)