Choose how you want to work with the agent
A standalone app gives you a place for conversation, files and results. An editor puts the agent alongside a project's files. A terminal lets you direct it through typed commands and inspect the tools it runs.
These are ways to interact with an agent. The model, available tools and execution location are separate choices. A desktop interface can control an agent working in the cloud.
Match the interface to the work
For a conversation or a first draft, start with an interface that makes it easy to supply material and revise the answer. For work across project files, look for a way to inspect changes and check the result.
Try one small task in the interface. Check whether you can see which files it used, understand the actions it proposes and return to the work later. Choose on the basis of what you can direct and review.
Check what the interface can reach
The interface may expose file access, a browser or connections to other applications. Those capabilities depend on the particular setup. Check the permissions and approval controls before handing over a task.
Voice mode is another way to give instructions and think aloud. On-device and cloud agents explains where the work runs.
This is one part of the harness. For examples of terminal, editor and app interfaces, see AI Tools & Skills.