An agent permission boundary defines the maximum authority available to an AI agent. It can restrict applications, files, accounts, browsers, network services, credentials and changes to external systems.
Unattended agents need especially clear boundaries because nobody may be present to interrupt an unexpected action. Removing unnecessary stored access and limiting reachable systems reduces exposure.
ELI5
An agent permission boundary is the line between what an AI agent may and may not do. It can allow safe reading or drafting while blocking sensitive changes, payments, messages, or access to unrelated information.
For example, an expense agent might be allowed to read receipts and prepare a reimbursement form, but need a manager's approval before sending money. The boundary should be enforced by the tools and system, not merely trusted as a polite instruction.
What can an agent permission boundary cover?
It can cover files, applications, accounts, tools, network destinations, credentials and the scope or duration of actions.
Why remove unnecessary access before an unattended run?
A narrower environment limits the damage from mistakes, unexpected navigation or instructions encountered during the workflow.
