I’ve run into more pitfalls than I expected when granting agents permissions.
The most dangerous thing isn’t that they do something wrong—it’s that they do the right thing, but you didn’t expect it. For example, they might delete a file they consider “unnecessary” while trying to complete a task.
So now I only grant three types of permissions: read access, the ability to create drafts, and revocable write access. There’s always a human-approval gate before anything is actually “executed.”
The limits of automation should be determined by the worst-case scenario, not the best-case one.
The most dangerous thing isn’t that they do something wrong—it’s that they do the right thing, but you didn’t expect it. For example, they might delete a file they consider “unnecessary” while trying to complete a task.
So now I only grant three types of permissions: read access, the ability to create drafts, and revocable write access. There’s always a human-approval gate before anything is actually “executed.”
The limits of automation should be determined by the worst-case scenario, not the best-case one.