Model versus agent
An AI model takes a request and produces a response, such as an answer or a draft. By itself, it cannot open your calendar, look up an order, or send a message.
An agent is software built around a model. It asks the model what step to take, uses a connected tool to do it, gives the result back to the model, and may repeat. The software controls which tools are available and what they are allowed to do.
People sometimes call this surrounding setup an AI harness, especially when it also supplies context and checks the results.
Two examples
- Parcel question: A model can suggest how to check delivery status. A support agent with permission can look up your order, check the courier’s update, and draft a reply based on what it finds.
- Meeting request: A model can suggest an agenda. An agent with calendar access can check free times, propose a slot, and create the event after you approve.
Keeping actions in check
An agent can choose the wrong step, misread a result, or repeat an action. The app should check permissions for each tool call, limit time and spending, keep a record of what happened, and ask for approval before important changes.