articleAugust 2026· 5 min reading time

Which AI model is inside? That does not yet tell you what the AI can do

An AI application may use Claude, Gemini or a model from OpenAI. That still does not tell you everything the AI can do. The software around the model determines what information the AI can see, what "tools" it is given and what actions it is allowed to carry out.

Claude Code shows the difference clearly

Anyone who uses Claude knows the ordinary conversation. You ask a question and Claude answers.

Claude Code also uses a Claude model as the thinking part. Around that model sits software that gives it far more scope. Claude Code can search and read files, change files, run commands and look at the result.

Suppose you ask:

Find out why this website returns an error and fix the problem.

The Claude model works out where to look. The software around it opens the files and runs the commands. The model gets the results back and then decides what the next step should be.

That software layer around the AI model is what we call an agentic harness.

You will also come across the terms agent harness and, in short, harness.

Codex from OpenAI works on the same principle: an AI model is the thinking part and the software around it gives that model the means to actually carry something out.

When does something become an AI agent?

Through such a harness, the AI can carry out several steps in a row.

It can look something up, carry out an action, look at the result and then decide for itself what the next step should be.

That is when we speak of an AI agent.

So an agent does not only tell you what you could do. Within the permissions it has been given, it can also carry out actions.

That is exactly why the software around the model matters so much.

1. The harness gives the AI "tools"

Suppose you tell an AI:

Change this customer's address.

The AI model can understand perfectly well what you mean. Without access to the customer system it cannot change anything.

If the agentic harness gives access to that customer system, the AI may be able to look up the customer and change the address.

If it also gets access to email, it may then be able to send a confirmation.

So the same AI model can do entirely different things depending on the "tools" it is given.

That can be access to:

•files;

•business software;

•databases;

•the internet;

•calendars;

•email;

•other programs.

2. The harness determines what information the AI gets

The harness can also supply information to the model at the moment it is needed.

For a question about an order, the right customer file can be retrieved first. The model then only sees the relevant data.

Rules set by the organisation can be passed along as well:

Discounts above 10 percent must be approved first.

Or:

This data may not be shared with external services.

The model decides what a logical next step is. The information and rules it receives help determine the decision it makes.

3. The harness lets the AI carry on

An ordinary chat usually works like this:

question → answer

An AI agent can carry on:

decide the step → carry it out → look at the result → decide the next step

That can happen several times in a row.

A large part of the power of AI agents lies there. You do not have to give a new instruction after every step.

Part of the risk lies there too.

A wrong answer is not the same as a wrong action

If a chatbot gives a wrong answer, you can read it and ignore it.

If an AI agent carries out a wrong action, the address may already have been changed, the email already sent or the appointment already booked.

That is why, with AI agents, you need to know more than how good the model is.

You also need to know:

•What information can the AI see?

•Which programs can it use?

•Can it only read, or change things as well?

•Which actions may it carry out on its own?

•When does a human have to approve first?

•How is it checked what the AI has done?

These are not technical details. They determine how much responsibility you are actually handing over to AI.

So ask more than: which model is inside?

A supplier may tell you that their application uses a well-known and powerful AI model.

Fine.

Then also ask:

What can that AI do with our data, our files and our programs?

And:

Which actions may it carry out without an employee giving permission first?

Then you have a far better idea of what you are really bringing into your organisation.

That is also why AIAdopt distinguishes between working with AI in the ordinary way and working with AI agents.

AI Agents teaches employees and managers to recognise when AI acts on its own, what they are handing over, where control is needed and when they have to step in.

Do you want to know first which AI applications are being used within your organisation and where the main risks lie? The AI Adoption Assessment then maps current AI use, shadow AI and the main points of attention.

The context of use also matters for AI literacy under Article 4 of the EU AI Act. Someone who only has texts improved works in a different context from someone who gives an AI agent permission to carry out actions on its own.

If you want to read more about the risks of AI that acts on its own, see our Insight One safe AI agent does not make a safe AI system as well.

Staying in control of what AI may do on its own

AI agents can take a lot of work off your hands. That makes it all the more important that employees understand what the agent can see, what it may do and where human control remains necessary.

Take a look at AI Agents, look at the AI Adoption Assessment or get in touch with AIAdopt if you want to work out first what makes sense for your organisation.