GitHub announced on October 26, 2026 that Copilot’s assisted permissions are now generally available. The feature adds a middle layer between letting an agent run every action automatically and requiring manual approval for all of them: a model evaluates each Copilot request, whether it’s running a command, accessing a file system path, or opening a URL, and only lets through without confirmation what it judges clearly safe.
An automatic arbiter between direct execution and manual review
Until now, teams running Copilot in agent mode had to pick between two extremes: approving every action one by one, which slows down any session that runs long, or allowing every command to execute automatically, which removes the safety net entirely. Assisted permissions add a third mode: requests classified as low risk run without interruption, everything else still comes back to the developer for approval. GitHub doesn’t publicly detail the exact classification criteria and points instead to its documentation on allowing safe commands.
Three permission modes, one new arbiter
| Mode | Behavior | Level of control |
|---|---|---|
| Manual | Every action requires approval | Maximum, but slows down long sessions |
| Assisted (new, GA) | The model filters low-risk actions, the rest still needs approval | Intermediate, arbitrated automatically |
| Allow all | Every action runs without approval | Minimal, can be blocked by admins |
A request judged clearly safe runs automatically; everything else comes back to the developer for approval.
Turning on assisted mode
The mode is configured directly in each Copilot surface, with no hidden global setting to track down.
# Copilot CLI: open the permissions picker
/permissions
# then choose "assisted"
# Copilot app: mode picker -> Tool permissions -> Assisted
# VS Code: permissions picker -> Assisted permissions
On the admin side, organizations keep control: it’s possible to block “Allow all” while keeping assisted mode available, to disable assisted mode separately, and to set the default mode for new sessions. If the default mode is unavailable or blocked by policy, new sessions start in manual approval.
What this changes for teams
The feature follows the same logic as the agent skills that already structure coding agents’ autonomy: give the agent more operational latitude while keeping a human checkpoint on what actually matters. For teams piloting agents from a terminal, the main gain is time, fewer interruptions for low-stakes actions, without giving up control over sensitive ones like deleting a file or reaching out to an unknown domain.
Key takeaway
Assisted permissions shift the balance between speed and control without removing it: a model does the first pass, a human keeps the final call on anything outside the obvious case. General availability covers Copilot CLI, the Copilot app, and Visual Studio Code; GitHub’s announcement doesn’t explicitly mention the autonomous coding agent or other IDEs.
Assisted mode is exactly what was missing to run an agent with confidence on a long session: manual mode burns through patience within minutes, and allow-all is unsettling the moment you work on a repo with secrets or deployment scripts nearby. What remains to be seen, once the feature is in hand, is whether the “low risk” classification holds up on ambiguous commands like a git push --force on a poorly named branch: that’s exactly the kind of case that separates a good heuristic from a bad one. — Simon Janvier
Further reading: the full announcement is on the GitHub changelog.
