On 26 October 2026 GitHub made assisted permissions generally available in the GitHub Copilot CLI, the GitHub Copilot app and Visual Studio Code.
What changed
When a Copilot agent wants to run a command, reach a file path or open a URL, a large language model now evaluates the request. Anything judged “clearly safe” goes ahead automatically. Everything else waits for a person to approve it. It is switched on per surface: /permissions in the CLI, the Tool permissions picker in the app, the permissions picker in VS Code.
For enterprises there are controls. Administrators can block the “Allow All” mode while leaving Assisted available, disable Assisted entirely, or set the default permission mode for new sessions. If the default is prohibited by policy, sessions start with manual approval.
The interpretation
Until now the choice was blunt: approve every action by hand, or let the agent run with everything allowed. Assisted permissions is a middle setting, and it is a notable one: the thing deciding what counts as safe is itself an AI model. That makes agents faster to work with, and it moves the safety judgement from a human click to a model’s assessment.
That is a reasonable trade for many teams. It is still a trade, and GitHub’s page does not publish an error rate for the safety judgement.
So what
If you commission software from a team using AI coding agents, ask which permission mode they run in, who set it, and whether destructive or production-touching actions always need a person. A supplier who can answer in one sentence has a process; one who cannot is probably running on defaults. That is the difference we mean by well-run AI-assisted development. If you want a build process with those controls written down, start a project.