Permission Request
The agent pausing to ask before doing something risky.
A prompt to a human before a sensitive action — writing outside the project, running a destructive command, sending a network request. The agent proposes; the person decides.
Well-designed permission requests are specific and rare. If users approve everything without reading, the prompt has become theatre and no longer provides safety.
Design approvals around reversibility and blast radius rather than around a list of scary-sounding commands. Writing inside a scratch directory is routine; force-pushing to a shared branch is not. Getting that calibration right is what keeps people from disabling approvals entirely.
Approval fatigue
The agent asks before every single file write.
People click approve without reading, and the control has become theatre. Scope the asks to external effects, credentials and destructive operations, and auto-allow the reversible work inside the project directory.
- docsComputer use tool ↗
Screenshot, mouse and keyboard as a tool loop, with the safeguards that belong around it: isolated environments, minimal privileges, network limits, human confirmation.
- docsConfigure permissions ↗
Allow, ask and deny rules for edits, shell commands and other tools: what a permission mode looks like when it is written down.
Permission Mode · Autonomy level
How much the agent may do without asking.
Human in the Loop · HITL
A person reviews or approves before the work lands.
Sandbox
An isolated environment where an agent can act without risk to real systems.
Guardrails
Constraints on what the system will do, enforced in code.