# Asking an agent to do a job still leaves the question of what it may do

> The same assistant can need different limits for different people, even when they ask for the same result.

Published 2026-11-08

Canonical: https://machines-on-machines.anth.us/posts/agent-permissions-user-controls/

Two people can ask an assistant to help with email and mean different things. One wants drafts to review. Another is comfortable with automatic replies to familiar contacts. The task description gives the agent a goal. A permission system determines which actions it may take to get there.

::image{
  src="assets/agent-permissions-user-controls/permissions.png"
  layout="inline"
  alt="A Claude settings screen shows modes governing when the agent may act and when it needs approval."
  caption="A Claude permission screen reproduced by the authors illustrates where a user can choose how tightly the system checks agent actions."
  credit="Alexandra E. Michael and Franziska Roesner, arXiv:2607.13718."
}

[Image source](https://arxiv.org/html/2607.13718v3)

Alexandra Michael and Franziska Roesner surveyed 21 proposals for user-level agent permissions and compared them with five commercial agents. Their framework separates how permissions are written or inferred from how the system enforces them. The researchers found gaps in transparency, usability and user control.

That distinction makes a product claim easier to examine. A setting can be easy to describe in the interface and still be difficult to enforce once an agent starts using tools. The survey gives readers a vocabulary for asking about both parts.

[Read the source](https://arxiv.org/abs/2607.13718).
<!-- papyrus-source-version: 3 -->