In VS Code, what do Copilot's agent permission levels — Manual permissions, Assisted permissions and Allow all — each change, and which is the default?
answer
- chosen per chat session
- the dropdown in the chat input
- one level defers to your approval settings
- an LLM judge in the middle
- Autopilot is a mode on the Agent Host
basics
~20 sManual permissions, the default, runs only what your tool, URL and terminal approval settings already allow and asks about the rest; Assisted permissions lets an LLM judge approve calls; Allow all runs every tool call without asking.
solid answer
~40 sPermission levels are chosen per chat session from the permissions dropdown in the chat input, and new sessions start at `chat.permissions.default` (Experimental), whose default is Manual permissions. **Manual permissions** applies your approval settings — saved tool approvals, `chat.tools.terminal.autoApprove`, `chat.tools.urls.autoApprove` — and prompts for everything else. **Assisted permissions** sends each tool call to an LLM judge and prompts only for calls it does not approve; it exists only for supported Agent Host sessions and is hidden unless `chat.assistedPermissions.enabled` is on. **Allow all** skips every confirmation. **Autopilot** also auto-approves all tools, retries on errors and answers its own blocking questions; on the Agent Host it is an agent mode, on the extension host a permission level. None of these turns the terminal sandbox off.
go deeper
Name the three levels and the default: Manual permissions. Know the picker sits in the chat input and that the level applies per session.
Explain what Manual permissions defers to, what the LLM judge in Assisted permissions does, and why Autopilot is a mode on the Agent Host but a level elsewhere.
Show that the level and the sandbox are separate layers, that worktree sessions are always Allow all, and when you would combine Allow all with an enabled sandbox.
Frame the team default: chat.permissions.default and managed policy against developer autonomy, and when disabling global auto-approval is worth the interruptions.
## Where the level is set Copilot's agent in VS Code can edit files, run terminal commands and call MCP or extension tools. Two layers govern those actions: **approvals** decide whether an action runs automatically or waits for you, and **sandboxing** limits what an approved terminal command can reach. The **permission level** is the session-wide switch on the approval layer. You pick it from the permissions dropdown in the chat input and can change it at any time during a session. New sessions start at the level in `chat.permissions.default`, an Experimental setting whose values are `default` (Manual permissions), `autoApprove` (Allow all) and `autopilot`; its default value is `default`. If an enterprise policy disables auto-approval, new sessions use Manual permissions regardless. ## The levels | Level | What it does | Notes | |---|---|---| | **Manual permissions** (default) | Runs actions your approval settings already allow; asks for everything else | Common read-only terminal commands run without asking; risky ones such as `rm` ask | | **Assisted permissions** | An LLM judge assesses each tool call; calls it does not approve ask you | Only for supported Agent Host sessions; shown when `chat.assistedPermissions.enabled` is on; off by default in Stable | | **Allow all** | Runs all tool calls without confirmation | A warning dialog appears the first time you select it | The first time you pick Assisted permissions, Allow all or Autopilot, VS Code shows a warning dialog. The docs are explicit that the judge behind Assisted permissions *can make mistakes* — it reduces interruptions, it does not replace your judgment. Manual permissions is only as strict as the settings behind it: - **Tool approvals** you saved earlier (for a session, a workspace or all future invocations), managed with **Chat: Manage Tool Approval** and cleared with **Chat: Reset Tool Confirmations**. - **Terminal rules** in `chat.tools.terminal.autoApprove`. - **URL rules** in `chat.tools.urls.autoApprove`, where contacting a URL and admitting its response are approved separately. - **Sensitive-file rules** in `chat.tools.edits.autoApprove`, which make edits to matching files wait for your approval. ## Autopilot: a mode or a level **Autopilot** lets the agent work until it judges the task complete. It auto-approves all tools like Allow all, retries when it hits errors, and responds automatically to questions that would otherwise block it. Where it lives depends on the session: 1. In a **Copilot** session — the Copilot harness running on the **Agent Host** — Autopilot is an **agent mode**, picked from the mode picker, separate from the permission level. 2. In a session on the **extension host**, such as a **Local** session, Autopilot is a **permission level**. Advanced Autopilot (Preview), enabled with `chat.autopilot.advanced.enabled`, hands the "is it done?" decision to a separate small model after each turn. Autopilot consumes AI credits like interactive chat. ## What the level does not decide - **The sandbox.** Sandboxing is independent of the level. An enabled sandbox still restricts terminal file-system and network access under Allow all and Autopilot. - **Worktree sessions.** A Copilot session started in a new Git worktree from the Agents window uses Allow all and cannot be changed; choose **Folder** isolation when you want approval prompts. A worktree isolates code changes but is not a security boundary. - **Global auto-approval.** `chat.tools.global.autoApprove` (default `false`) removes prompts in every workspace. The session commands `/yolo` or `/autoApprove` bypass approvals only for the current supported session, and `/disableYolo` or `/disableAutoApprove` restore its default level. - **Which tools exist.** Enabling a tool makes it available; it never bypasses approval. ## Which levels a session offers Not every session shows every level: - A **Copilot** session with **Folder** isolation offers Manual permissions or Allow all, plus Assisted permissions once `chat.assistedPermissions.enabled` is on. - A **Copilot** session in a **New Worktree** is fixed at Allow all. - An organization can hide Assisted permissions by disabling global auto-approval, and when policy disables auto-approval, new sessions start in Manual permissions whatever `chat.permissions.default` says. If a teammate reports that the agent "never asks", the first two questions are which level the session shows in the dropdown and whether it is a worktree session. ## Picking a level for a session Manual permissions fits work in a repository you have not reviewed, or whenever the agent may touch credentials, deployment scripts or shared infrastructure. Assisted permissions, where available, trims prompts for routine calls while keeping a human on the unusual ones. Allow all and Autopilot belong to workspaces you trust — ideally with the terminal sandbox on, because then the prompts disappear but the operating-system boundary stays.
- Why can't you choose Manual permissions for a Copilot session that VS Code's Agents window started in a new worktree?Worktree sessions use **Allow all** and the level cannot be changed, because their edits are kept apart from your active workspace. If you want approval prompts, start the session with **Folder** isolation. The worktree only separates code changes; it does not restrict commands or network access, so it is not a security boundary.
- How do Allow all, `/yolo` and `chat.tools.global.autoApprove` differ in scope?Allow all is the session's level, picked from the dropdown. `/yolo` or `/autoApprove` bypass approvals for the current supported local or Copilot CLI session, and `/disableYolo` or `/disableAutoApprove` restore the session's default level. `chat.tools.global.autoApprove`, off by default, auto-approves tools in every workspace.
- Does switching a Copilot session to Allow all turn the terminal sandbox off?No. Sandboxing is independent of the permission level: Allow all and Autopilot skip approval prompts, but an enabled sandbox still restricts terminal commands' file-system and network access. Because the sandbox covers only terminal commands, other actions rely on tool and URL approvals under Manual permissions.
- Allow all is the default permission level for new agent sessions.
- Assisted permissions means a teammate or admin reviews each tool call.
- Choosing Allow all also lifts the terminal sandbox.
- Autopilot is simply another name for Allow all.
- Manual permissions prompts before every tool call, including read-only commands.