How Agent works
See how one natural-language request becomes useful work across online tools and your Apple device.
Agent follows the same request and action loop whether you start from the Shortcuts app, Siri, the Share Sheet, an Action Button, or a Home Screen widget.
1. You start with an intent
Tell Agent the result you want in ordinary language. You can begin with only a request, or include something you already have open, such as text, a link, an image, media, or a PDF.
For example:
Add this event to my calendar and tell Jamie where it is.
If the request begins in the Share Sheet, the shared item travels with it as context. Agent also receives basic device context supplied by the Shortcut, such as the current time and the app where the request began.
2. The brain understands and plans
The selected AI model reads the request and available context. It identifies the desired outcome, extracts useful details, and decides which steps are needed.
If an important detail is missing, Agent can ask a follow-up question instead of guessing. A request can also take several steps: Agent may use one result to decide what should happen next.
3. The hands use tools
Tools let Agent move from understanding to action. They fall into two practical groups:
Online tools
An online tool runs as part of the Shortcuts AI service. For example, web search can find current information that was not included in the original request.
Actions on your device
For an Apple app or system action, the service sends an instruction back to the Agent shortcut. Apple Shortcuts performs the supported action on your device. It might add a Calendar event, create a Note, open Maps, or prepare a Message.
Your device may request permission the first time a Shortcut uses an app or system feature.
4. Results can continue the loop
Some actions simply finish the task. Others return information to Agent so it can continue planning. This is how one request can become a short sequence instead of a single isolated command.
For the event example, Agent could:
- Read the shared flyer.
- Identify the event name, time, and location.
- Look up a missing venue detail if needed.
- Add the event through Calendar on your device.
- Prepare a Message for Jamie.
- Confirm what was completed.
You begin with the outcome you want. The finished work appears in the app where you need it.
Agent does not automatically read everything in your apps. It works with the input you submit, the context sent for that request, and the results returned by enabled tools. Read Architecture and data flow for the full boundary between the device and the online service.
Control what Agent can do
The tools available to Agent are controlled from Settings. You can disable tools you do not want used and add instructions that Agent should follow on every run.