Agent mode is what lets LiveAI read, write, or modify files on your behalf — scaffolding a project, fixing a bug, reorganizing a folder. By nature, it's also the most sensitive feature in the app: you're giving a language model the ability to act, not just to respond.
Four levels, not a single switch
Agent mode isn't a simple on/off toggle. Four permission levels are available under Settings → Permissions:
- Disabled — the agent doesn't touch anything, plain chat only.
- Read-only — the agent can view files but never create, modify, or delete them.
- Confirm every action — every write waits for your explicit approval before it runs.
- Full access — the agent acts continuously, for long tasks where confirming every step would slow things down without adding real safety.
Review before applying
Even in full-access mode, you can turn on review before applying: every file change shows its exact diff (what changes, line by line) and waits for your approval before touching disk. Nothing is ever written blindly.
The shell is a special case
The agent also has full shell access (PowerShell/bash), which isn't confined to the chosen working folder — unlike the file tools. For that exact reason, a confirmation is always required before every shell command, no matter which permission level is otherwise selected.
Undo, even after the fact
After a task that has modified files, a button lets you undo everything at once. You can also drop a checkpoint partway through a task to roll back to that exact point later, without having to undo everything if only the tail end of the task went wrong.
The general principle: the harder an action is to undo, or the broader its scope (shell, deletion), the more systematically explicit confirmation is required — regardless of the overall permission level chosen.