Allow, ask, block
A permission model a non-engineer can explain in a meeting, and an agent cannot talk its way around.
Nimind Team
Security
Most agent failures are not model failures. They are authority failures: the agent did something nobody explicitly allowed, because nothing was there to stop it. A prompt is not a permission system. It is a suggestion with good manners.
Three answers, not two
- Allow — reversible and inside the job. Reading the repo, drafting, running tests, opening a pull request.
- Ask — consequential or public. Sending mail, spending money, merging, deleting a project.
- Block — anything the role should never do, no matter how the request is phrased.
The middle answer is the one that makes the model usable. With only allow and deny, teams allow too much, because denying everything produces an agent that cannot do the job and a team that turns the checks off.
Rules belong to the role, not the prompt
Permissions live on the agent file, not in the conversation. A prompt can be argued with by a clever request. A rule is evaluated before the tool runs, outside the model’s reach.
Write the list before you need it
- 01
Name the five actions that would embarrass you
If one of them happened without review, who would have to apologise to whom? Those go in Ask.
- 02
Name the three that would be worse
Deleting a database, posting publicly, moving money to a new recipient. Those go in Block.
- 03
Allow everything reversible
The rest of the job should not need a human in the loop, or the human becomes the bottleneck the agent was meant to remove.
Expect the list to change. The first week of real work surfaces actions nobody thought of, which is exactly why the rules live in version control and not in a settings page someone opened once.
Published September 24, 2026 · Nimind Team
Back to all posts
