# Meet models and agents

> Explain how a model becomes part of a system that can act.

[Canonical HTML page](https://trustcyber.ca/research/allow-hold-deny/lessons/01-ai-and-agents/)

- Author: Junior Williams
- Type: Research companion addition

## Learning objective

Explain how a model becomes part of a system that can act.

- [Original paper: Introduction · foundations added by this companion · page 1](https://trustcyber.ca/research/allow-hold-deny/original.pdf#page=1)
Source pages: 1
## A model produces an answer

An AI model is software trained on examples to produce an output from an input. The input might be a question, a document or a description of a situation. An output might be text or a selection from a small set of choices. Training does not make every output correct.

An **agent** combines a model with software that can pursue a task and use tools. A tool might read a file or send a message. The surrounding software decides which tools exist and when an action is carried out.

## Follow one proposed action

A student asks an assistant to prepare a newsletter. Drafting the text produces a document. Sending it changes something outside that document: recipients receive a message. An authorization decision asks whether that proposed send is permitted under the supplied rule and facts.

This paper tests the decision step. It supplies written situations to four model configurations and records their answers. It does not connect those answers to a tool that sends the newsletter.

<details><summary>Go deeper: model, service and configuration</summary><p>A model is one component. A service wraps it in an interface. A tested configuration includes the model, interface, software settings and execution conditions. A result for one configuration does not identify the effect of each component separately.</p></details>

## Explore the idea

Select a step in the newsletter example and inspect its role.
- Draft the message: The system prepares text. The newsletter has not been sent.
- Decide whether sending is permitted: The decision compares the proposed send with a rule and supplied facts.
- Send the message: An execution tool would contact recipients. This experiment did not execute that step.

## Interactive activities

Use the canonical HTML page for controls, live results and reset buttons.

## Check your understanding (optional)

What was actually executed in this study?
- Requests asking models to classify written situations: Correct. The experiment recorded decisions about proposed actions. The actions themselves were not performed.
- The actions described in every case: The paper explicitly says no proposed action was executed.
- Only a human survey: A human reviewed the reference key, and four configurations supplied model responses.

## Take this forward

Next, separate the ability to carry out an action from permission to carry it out.

## Reference material


## Lesson navigation

- [Next: What is the paper asking?](https://trustcyber.ca/research/allow-hold-deny/lessons/02-the-research-question/index.md)

## Companion navigation

- [Overview](https://trustcyber.ca/research/allow-hold-deny/index.md)
- [Both authorization studies](https://trustcyber.ca/research/authorization/index.md)