Check where the tools and files live
An on-device agent runs tools against files on your computer. A provider-hosted cloud agent works on a computer the provider manages. A self-hosted agent runs on a computer or server you manage.
The interface can be elsewhere. You can direct a cloud agent from a desktop app or a browser.
Check where the model runs
An agent working on your computer can call a hosted model. The instructions, evidence and tool results included in those calls go to that model's service. Running tools locally therefore leaves a separate question about where the model processes their results.
Local model inference runs the model on your hardware. A self-hosted agent can still call a hosted model. Describe tool execution and model inference separately when choosing or explaining a setup.
Match the location to the work
Work that depends on files and applications on your computer needs a way to reach them. Work that continues while your laptop is closed needs an execution environment that remains available, with access to the required material.
With a provider-hosted environment, inspect the available tools, file access and approval controls. With a self-hosted environment, include the work of maintaining the machine and checking its scheduled output.
Execution is one part of the harness. The interfaces guide covers how you direct it. The open-weight models guide covers a separate model deployment choice.
These distinctions follow the interfaces and execution explanation in AI Tools & Skills.