VS Code 1.143 macht Copilot Harness zum Standard

You open VS Code after an update and something feels different. A new picker here, a new setting there. You wonder: "Did they move my stuff again?" 😅

VS Code 1.143 has a few changes like that. Some are big, some are small, and several are still experimental. This post walks through the highlights in plain English so you know what changed and what to try first.

VS Code ships a new version regularly, and 1.143 is the latest release covered here. Think of it like a new season of your favorite show: same main characters, a few new storylines.

The theme of this release is AI agents. Agents are AI helpers that can work on coding tasks for you, and this release makes them easier to run, compare, review, and clean up after.

A quick note on the word "harness." It's the engine that runs your agent session. The Copilot harness is built on the Copilot SDK, a shared foundation meant to keep behavior and improvements consistent across Copilot products.

Using AI agents is fun until the questions start:

This release goes after those questions. It gives you ways to compare results, review changes, and clean up leftovers. The release notes also describe performance work for people who keep very long chats.

New sessions use the Copilot harness. You can keep these sessions running in the background and reach them from both the editor window and the Agents window. They run through the Agent Host.

Example: You start a task in the editor, then switch to the Agents window to check on it. The same session is there.

The release notes also mention reduced delays caused by MCP servers that don't finish loading their tools. If you ever waited a long time for a chat response, that fix is aimed at you.

You can run the same task with two or more models at the same time. Each attempt gets its own session, and you can mix Copilot, Claude, and Codex harnesses with different reasoning levels.

You can also pick an optional Judge model. It evaluates the attempts and shows a scorecard with time and token usage. If you like, you can select Synthesize to build on the recommended attempt using ideas from the others. Synthesis only starts when you ask for it.

Example: You need a form validation helper. Instead of guessing which model writes the cleanest version, you run several and compare side by side.

To try it, enable sessions.chat.compareAgents.enabled in the Agents window, then choose Compare Models in the model picker.

Worktrees let agents work on several features without stepping on each other. The catch is that inactive ones pile up and eat disk space.

Run Chat: Open Worktree Cleanup from the Command Palette to see how much space they use. You can filter by inactivity, review sizes, and pick what to remove. Cleaning up marks sessions as done and deletes their worktrees, and you can restore a session later to recreate its worktree. Active, running, needs-input, and pinned sessions are protected.

Example: It's like cleaning your desk. You keep the papers you're using and clear the ones from three weeks ago.

File-level Keep (Experimental). Enable chat.editor.workingSet.enabled and you get a working set above the chat input. Select Keep on a file to mark it as reviewed. The edits stay on disk, and the file comes back if the agent changes it again.

More Discard options in the Agents window. Discard Changes now works in more changesets, not just Uncommitted Changes.

Opening a long Agent Host chat used to freeze the window for a long time. In the release notes' measurements (median of three runs on macOS), one chat with about 1,150 requests went from 35.6 s to 9.0 s on a cold open, and the unresponsive time dropped from 28.3 s to 0.6 s.

The same button behaves differently depending on where you click it. This is worth knowing before you click:

If you only want to undo the agent's most recent attempt, Last Turn Changes is the narrowest option.

1. Mixing up Keep and Discard. Keep marks changes as reviewed and leaves them on disk. Discard reverts them. Read the label before you click. 👀

2. Forgetting a feature is experimental or preview. If something looks different from the docs, check whether it's a Preview or Experimental feature behind a setting.

3. Trying on-device models without the hardware. MAI Code 1.1 Flash requires GitHub sign-in and an NVIDIA GPU with 96 GB or more of GPU memory. You also need a compatible llama.cpp build, plus chat.copilot.localModel.llamaServerPath and chat.copilot.localModel.enabled set. Responses that run on your device are labeled On-device and don't use AI credits.

4. Ignoring cleanup until the disk is full. Worktrees add up quietly. A quick review is easier than an emergency one.

5. Skipping the enterprise notes. If you work in a company setup, admins can control things like managed defaults, sandbox bypass, and whether the Local harness is allowed. If a toggle is locked, policy is probably the reason.

VS Code 1.143 is a practical release for anyone working with AI agents. The Copilot harness is now the default, Compare Models helps you pick the best result, worktree cleanup keeps your disk tidy, and the new review tools make agent edits easier to manage. Several of these are still Preview or Experimental, so try them with curiosity and keep your expectations realistic. 🚀

If this helped, share it with a developer friend who's wondering what changed in VS Code 1.143. For more practical dev posts, visit hamidrazadev.com. And tell me in the comments: which feature are you trying first? 😊