Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Decide whether Copilot's assisted permissions should approve anything for you

Written November 1, 2026

For a Copilot user drowning in approval prompts: assisted permissions, generally available since 26 October 2026 in the Copilot CLI, the Copilot app, and Visual Studio Code, let a model approve the requests it judges clearly safe, and leave the rest to you. Decide per surface and per project — and read the conflict over enterprise "ask" rules below before treating an ask rule as a human checkpoint.

What changed

Assisted permissions are now generally available in GitHub Copilot CLI, GitHub Copilot app, and Visual Studio Code. When Copilot requests permission to run a command, access a file system path, or open a URL, a large language model evaluates the request. Requests classified as clearly safe automatically proceed; other requests require your approval. Enterprise administrators can use enterprise-managed settings to block Allow All while keeping Assisted available, or to disable Assisted separately, and a default permission mode can be chosen for new sessions — if the default is unavailable or prohibited by policy, the session starts with manual approval. The announcement names only those three clients. It does not name JetBrains, Visual Studio, or the Copilot cloud agent, so do not assume the mode exists there.

GitHub's documentation, Allowing safe commands to run without manual approval, calls assisted permissions a middle ground between manual approval and granting broad permissions with an allow-all option. Existing deny rules and enterprise restrictions still apply, and requests already allowed by your permissions can run without a prompt or a model review. Existing rules and human approvals are evaluated before a model review. Copilot selects the model used for the safety review, which can differ from the model you selected for your conversation; if you use your own model provider, the review uses your session's active model. GitHub publishes no rule set for what counts as clearly safe, and no rate for how often the model's judgment is wrong. Treat assisted mode as a judgment call by a model, not as a policy with stated rules.

The three modes, side by side

Assisted permissions do not review everything. They do not review other tool calls that enterprise rules require you to approve, or requests to bypass local sandboxing. Those requests remain subject to the session's normal permission handling. If a safety review is unavailable or cannot evaluate a request, the request follows the session's normal permission handling.

Enable it per client — the scope differs

Each client has its own route, and in Visual Studio Code the scope is narrower than the name suggests. In the Copilot CLI, run /permissions and select assisted; /permissions assisted turns it on for the current session, /permissions manual returns to manual approval, and /permissions show reports the current mode. To make assisted the starting mode for new interactive sessions, set defaultPermissionMode to "assisted" in your user configuration file, ~/.copilot/settings.json. If your administrator sets a managed default, that default takes precedence, and you can still choose a different mode for a session if policy permits it. Resumed sessions keep their saved permission mode, subject to current policy restrictions. In the Copilot app, open the mode picker, click Tool permissions, then select Assisted; the selection applies to the current local session and is saved as the default for new local sessions. Tool permission controls are not available for cloud sessions. In Visual Studio Code, open the permissions picker and select Assisted permissions — but only in a Copilot harness session with Folder isolation. Microsoft's approvals documentation states that the option is not available for Local harness sessions, and that Worktree sessions always use Allow all. Check which harness and isolation your session uses before assuming assisted mode is in effect.

The ask-rule conflict, stated not resolved

GitHub's own documents disagree about what an enterprise "ask" rule means once assisted permissions are on, and this page does not resolve the disagreement. The assisted-permissions guide states: "In Copilot CLI, enterprise ask rules for shell commands, file reads, file writes, and URLs can be satisfied by model approval when the request is eligible for review. This means those rules do not always result in a prompt for your confirmation." The enterprise managed-settings reference states the opposite for ask rules generally: an ask rule "requires fresh, one-time approval before a specific operation can proceed," and "a managed ask rule can't be satisfied by bypass mode (also known as allow-all or YOLO mode), an auto-approval setting, a hook or other approval shortcut, or a grant persisted from an earlier approval. The same operation prompts again the next time it's requested." Both wordings are GitHub's, in documents that cover different scopes — the first names the Copilot CLI and eligible requests, the second states a general rule for managed settings. Neither document, as read, explains how the two fit together. Until you have tested the exact client and operation in your own environment, do not treat an ask rule as a guaranteed human checkpoint, and do not treat model approval as equivalent to that checkpoint either. Enterprise deny rules still apply under both wordings, and that part is not in conflict.

Steps: decide, then test the decision

  1. Start from manual on unfamiliar or high-stakes work. A new repository, a task that deletes, publishes, or changes anything outside the project, and any work with credentials nearby all stay on manual approval. Assisted mode earns its place on bounded work in a familiar project, where changes are reviewed before they land.
  2. Turn assisted on for one session, on one project. Use the per-client route above for the session in front of you, not the default for every future session. Work a normal task and count what it approves without prompting you. If you cannot say, for each auto-approved action, why it was safe, the mode is approving more than you can follow — switch back with /permissions manual or the picker.
  3. Check the mode before trusting it. In the CLI, run /permissions show. In Visual Studio Code, confirm the harness and isolation: Folder isolation in a Copilot harness session is where assisted mode applies. If the session is a Local harness session or a Worktree session, assisted mode is not doing what you think it is.
  4. Decide the default only after the test. If the one-session test went well, set defaultPermissionMode to "assisted" for new interactive sessions, knowing an administrator's managed default takes precedence and a blocked default falls back to manual approval. Keep the default at manual for any machine or project where the test has not been run.
  5. Keep sandboxing and review in place, whichever mode you pick. GitHub warns that the model can make mistakes and approve dangerous or unintended actions, and that assisted permissions are not a substitute for sandboxing or access restrictions. Run the session sandboxed or otherwise isolated where you can, and review the actions Copilot takes as you go. Assisted mode reduces prompts; it does not remove responsibility for what Copilot does.

Non-interactive runs behave differently

In autopilot mode and non-interactive runs, requests that still require approval are denied instead of prompting anyone. When the safety review requires approval, its explanation is passed back to Copilot so it can try a different approach. For a non-interactive CLI run, use the --assisted-approval option, or set the COPILOT_ASSISTED_APPROVAL environment variable to true — for example: copilot -p "Fix the failing tests" --assisted-approval. The option is only for non-interactive runs; using it for an interactive session makes the CLI exit with an error. Two traps deserve a direct check. First, when successfully enabled, --assisted-approval takes precedence over --allow-all-tools, but it does not change --allow-all-paths or --allow-all-urls — combining it with --allow-all can still leave path and URL access unrestricted. Second, if assisted permissions cannot be turned on because enterprise settings block it, the CLI writes a warning to standard error and does not cancel a separate --allow-all-tools option you supplied, so tool requests may still be approved without a safety review. Read the output of automated runs for that warning; an unattended run that silently fell back to allow-all is worse than one that stopped.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Review a pull request with a coding agent · Run a reasoning task on a schedule with GitHub Agentic Workflows · Run work that spans several repos in one Copilot Project