A laptop displays a coding interface and AI assistant beneath a banner announcing “1.147 Stable” agent-focused updates.
VS Code 1.147 is out, and it is mostly an agent release. Visual Studio Code 1.147 shipped on November 18, 2026. The release metadata lists it as the Stable edition, with download version 1.147.0. The headline you may have seen carries an "(Insiders)" tag. That label comes from the Insiders-build coverage, and the release page itself describes a Stable release. This article treats it as 1.147 Stable.

Most of the new features sit in the Agents window and the Copilot harness. Many are labeled Experimental or Preview, so check each feature's stage before you rely on it. Several of the more practical changes are aimed at remote development and enterprise administrators.

A laptop displays a coding interface and AI assistant beneath a banner announcing “1.147 Stable” agent-focused updates. Scheduling and drafting agent requests​

Send Later (Experimental) postpones a chat request until a time you pick. It is controlled by sessions.sendLater.enabled and works in the Agents window only.

  • You schedule from the Send Later menu on the send button. Choose Date and Time, pick a day and time, then select Schedule.
  • A request scheduled from a new session waits in the Drafts section of the sessions list. A request for an existing chat waits in that chat's Scheduled area.
  • If the agent is still busy at the scheduled time, the request waits its turn. It does not interrupt.
  • You can also ask in plain language, for example "Check the build in 10 minutes." The agent schedules a one-time, text-only request in the current chat.
  • You can edit, send immediately, reschedule, or cancel a request from the hover options. A clock indicator marks sessions with pending requests, and a Scheduled Only filter finds them.

The limits matter here. Scheduled requests work only for sessions on This Computer, and they require chat.automations.enabled. The local agent host must be running at send time. Scheduling doesn't wake a sleeping machine. A missed request is processed once the host is available again. So a laptop that sleeps overnight will not run your "follow up in the morning" request until it wakes.

Drafts (Experimental), enabled with sessions.drafts.enabled, replaces the old single unsent-request box. Leaving a populated new session saves it to the Drafts section without sending anything. Each draft keeps its text, attachments, inline references, workspace, agent, model, and other configuration, and drafts survive reloads and restarts. A draft is deleted only after the request is accepted, so a failed send doesn't lose your work. Turning the setting off hides Drafts but keeps the saved ones.

Session management and persistence​

  • Share remotes as GitHub Environments: In the Remote Environments editor, you can share a connected SSH host, WSL distribution, or Dev Tunnel. That lets you pick up its agent sessions from another device or a supported Copilot app with your GitHub account. Selecting the action again stops sharing.
  • Dev Container persistence (Experimental): Rebuilding a Dev Container used to wipe the Copilot chat sessions stored inside it. The settings dev.containers.stateVolume.enabled and chat.agentHost.devContainer.stateVolume.enabled put user and agent state on a Docker volume. A container gets the volume only if it is created with the setting on, so existing containers need one rebuild first.
  • Inbox mode (Experimental): You can mark a session as acknowledged. It moves to an Acknowledged section without being marked read, done, archived, or closed. It returns to the Inbox when new activity arrives. Workspace and environment pickers now filter every section, including pinned and archived sessions, and the filter persists across reloads.

Model routers, comparisons and visual answers​

Model routers (Experimental) are enabled with "chat.copilot.modelRouter.enabled": true in user settings.json. A router is a YAML file in ~/.agents/model-routers/. A decision model, gpt-6-luna-decisions, matches each request to one of your routes. The routers appear in the model picker for Copilot harness sessions. You can author one by typing /model-router in a Copilot harness chat and describing what you want. The example in the notes sends multi-step coding work to Opus and quick questions to GPT-6 Luna. That is an illustration, not a recommended setup. Microsoft calls the feature early-stage and asks for feedback.

NTCompatible's coverage describes the pitch as routing tasks to cheaper or stronger models automatically. That is a reasonable reading. The notes themselves only say you describe which model suits which kind of work.

Model comparison (Preview) now keeps comparisons beneath the chat that started them, with an Open Comparison action. Hovering over a comparison shows how it was set up: repository, isolation, base branch, each attempt's model and effort level, judge, and synthesizer.

Visual responses are off by default, in the Agents window only.

  • chat.htmlPreview.enabled with the /html-preview skill renders HTML, CSS, tables, inline SVG, and expandable sections. Scripts, forms, external links, and external resources are not supported.
  • chat.nativeComponents.enabled (Experimental) with the /vscode-ui skill renders metrics, badges, tables, and bar charts in your theme, including high-contrast themes.
  • Right-click a visual and choose Show Source to see its markup. HTML previews can also be opened larger or saved as a self-contained HTML file.

Permissions, sandboxing and automations​

Assisted permissions: An LLM reviewer approves low-risk tool calls in Copilot harness sessions. The new chat.permissions.assisted.instructions setting lets you tell it what you consider risky. For example, you can require asking before package installs or pushes, and skip approval for tests and read-only Git commands.

  • Your instructions take precedence over the reviewer's defaults.
  • They apply to sessions that are already open.
  • They can't override permission rules, organization policy, or the sandbox, and they aren't sent to the agent.
  • Requests started by automations or other agents are no longer treated as if you wrote them. Expect more approval prompts in those sessions.
  • Your answers to agent questions and plan reviews now count, including refusals. A refusal can block an action that was approved earlier.

Sandbox retry: If a terminal command fails for lack of file or folder access, you can grant access for the session (the default), the workspace, or the user. With sandbox bypass enabled, Copilot then retries inside the sandbox. None of these options permits running outside the sandbox.

Automations get guided creation through an agent, a Create Manually option, and end dates in Days, Weeks, or Months. Automation runs are now hidden from session lists by default, and an Automations filter shows or hides them.

Browser Use and remote HTML previews​

Browser Use (Preview) lets agents work in your default Chrome or Edge. That includes reusing authenticated sessions, which is the useful part and also the part to treat carefully. The requirements are:

  1. The GitHub Copilot browser extension.
  2. A Copilot harness session running on your own machine.
  3. Node.js 22 or later on your PATH.

Setup steps:

  1. Enable chat.agentHost.browserUse.enabled and restart VS Code.
  2. Open a local Copilot session and run /plugin enable browser-use.
  3. Ask for a browser task, such as checking navigation on localhost:3000 and capturing a screenshot.
  4. Approve site access when prompted, and do sign-ins and sensitive steps yourself.
  5. If something fails, run /browser-use:doctor.

Remote HTML preview (Preview): Set workbench.browser.enableRemoteProxy. Then right-click an HTML file from an SSH, WSL, or Dev Container workspace and choose Open in Integrated Browser. The page keeps its file:// address, and its resources load from the remote workspace. Remote files load only from trusted folders.

Two smaller browser settings arrive as well. The integrated browser is now an option in workbench.externalUriOpeners, so you can send a pattern such as *.contoso.com to it. And workbench.browser.dimOnOverlay can be turned off to stop the page dimming when hovers overlap it. Notifications still dim it.

Enterprise controls​

Previously, any Copilot enterprise-managed setting disabled the Claude and Codex harnesses. Administrators can now list them in vscode.allowedHarnesses.

  • Valid entries are lowercase claude and codex.
  • An empty list blocks both. With no setting, both stay disabled as before.
  • It goes in file-based or MDM-managed settings. Server-managed settings don't support it yet.
  • Other controls still apply, such as vscode.temporaryAllowLocalHarness: false.

The caveat is that permitting a harness does not make it enforce your managed settings. Claude and Codex sessions run without Copilot's managed permission, sandbox, MCP, and plugin controls. Admins are trading governance for choice, and should do that deliberately.

Managed Shell(...), Bash(...), and PowerShell(...) rules in permissions.deny, permissions.ask, and permissions.allow now accept JavaScript regular expressions in /pattern/flags form. The documented example requires approval for commands starting with git push or git pull.

Editor, pull request and chat changes​

  • Markdown: Tables stay rendered while you edit them cell by cell. <details> and <summary> render as collapsible sections, expanded only when open is set.
  • Chat: There is an experimental Screenshot App Window option, with Ctrl+Shift+Alt+S on Windows and Linux. A Mark Up tool lets you crop, draw, or add text to image attachments. Inline attachment pills can now be dragged. They are on by default in Insiders and off in Stable.
  • Layout: workbench.sizeCustomizations sets panel corner radii from 0 to 8 pixels and spacing from 0 to 16. Details panel placement is experimental and applies to the Agents window.
  • Accessibility: High-contrast themes get dashed hover and solid focus outlines. Sessions awaiting input get clearer indicators.
  • GitHub Pull Requests extension (0.180.0): Adds an Approve workflows to run button for workflows awaiting maintainer approval. The overview goes single-column below 1000px instead of 768px. An experimental Agents-window tabbed view is also included.

Deprecations and removals​

  • VS Code desktop on macOS Ventura and Sonoma is deprecated.
  • VS Code desktop on 32-bit ARM Linux is deprecated.
  • The experimental Status Bar AI usage statistics (editor.aiStats.enabled) are removed because of low usage.
  • The Local agent harness is slated for removal in February 2027.

What to do about it​

Windows users get the same settings and shortcuts as everyone else. The WSL and Dev Container items will matter most to Windows developers working in those environments.

  1. Update normally through Check for Updates. Most of the new agent features do nothing until you enable them.
  2. Enable features one at a time. Many are Experimental or Preview, and Warp2Search's coverage notes that many sit behind experimental switches and advises caution in managed environments.
  3. Treat Browser Use as a security decision. An agent reusing your signed-in browser session is powerful, so keep the site-approval prompts on.
  4. Plan for the Local harness removal. Anyone depending on it has until February 2027.
  5. Admins should read the harness caveat twice. Allowing Claude or Codex lifts a block but adds no controls.

This review rests on the official release notes and on third-party coverage. It involved no hands-on testing, so behavior in your environment may differ.

 

References

  1. Visual Studio Code 1.147 (Insiders) Visual Studio Code 2026-11-18T17:00:00Z
  2. Visual Studio Code 1.147 Stable: Browser Use, Model Routers, and Send-Later Land in November Update warp2search.net
  3. vscode-docs/release-notes/v1_147.md at main · microsoft/vscode-docs · GitHub github.com