Permissions and Confirmations
Understand why reads, writes, execution, and external actions require different levels of authorization.
Why confirmations exist
The AI can propose or perform actions, but you remain the one granting authority. Reading data, writing files, running commands, sending messages, and deleting external data all carry different levels of impact, so they should not share a single vague master switch.
Common action levels
- Read: View files or connected-app data you have explicitly selected.
- Write: Create or modify local files and product records.
- Execute: Run commands, scripts, or agent tools.
- External changes: Send email, create events, or modify or delete remote data.
Deciding whether to approve
Verify the target, the scope, and the expected outcome. For sensitive actions, review the plan or diff first. Never approve commands, services, or target addresses you don't recognize.
Permission modes
Nexts may offer different agent permission modes, but client-side prompts only improve the experience — actual data and system permissions are still enforced by the local backend, the operating system, and external services together.