Permissions
Izin
The rules that decide what the agent is allowed to touch.
Permissions are the rules that decide which files an agent may read, which commands it may run, which systems it may write to. They are the enforcement point for everything the system prompt asks for, because a rule written in prose is only a suggestion to a model that is optimised to be helpful.
The principle to follow is least privilege: give the agent exactly what the task needs, and nothing more. An agent editing one report does not need write access to the whole repository. A summariser that only reads does not need the ability to run shell commands at all.
Permissions have to be enforceable, which means they belong in code or in the filesystem, not in a prompt. It also means they need a shape that survives contact with real use: a path allowlist, a network allowlist, a separate identity per integration, and a way to revoke quickly when something goes wrong.
Related terms
Guardrails
Checks that stop or constrain an action before it happens.
Approval
The moment a person says yes before the agent does something it cannot undo.
Human in the Loop
A person who reviews, approves, or corrects the agent's work at key points.
Agent
A model with tools, a system prompt, and a loop that lets it act and observe results.
Harness
Everything around the model that turns it into an agent: tools, system prompt, permissions, hooks.
Want this working inside your team?
It starts with one conversation about your workflow.