VS Code 1.143 Review: Still Free, Now Agent-First
Microsoft's code editor makes Copilot the default engine for AI sessions and lets you pit models against each other – but it is still a superb editor underneath.
Few programs have conquered developers' desktops as completely as Visual Studio Code. It is free, it starts quickly, it speaks practically every programming language through extensions, and it has become the default answer to "which editor should I install?" on Windows. The catch in 2026 is that the release notes read less like an editor changelog and more like a report from an AI lab. Version 1.143, which landed in the third week of October and has already received the 1.143.1 bugfix update, is a perfect example: most of its headline features concern agents, models and sessions. So we asked a simple question – is VS Code still a great editor for everybody, or only for people who want an AI co-worker?
First impressions and installation
On Windows you can choose between a User installer, which needs no administrator rights and installs into your profile, and a System installer for all accounts on the machine. Microsoft states a download of under 200 MB and an installed footprint of under 500 MB, and recommends a 1.6 GHz processor and 1 GB of RAM – modest numbers for a modern development tool. Officially, VS Code supports the 64-bit Windows client versions that Microsoft itself still services. There are no bundled offers, and VS Code itself costs nothing, for private or commercial use.
New in 1.143 is an extra option in the Windows installer: it offers to install the GitHub Copilot CLI alongside the editor (the option is hidden if you already have it, and organizations can switch it off). Worth knowing: the AI features require a GitHub sign-in and a Copilot plan. A Copilot Free plan exists, paid plans unlock more. If you do not want AI at all, VS Code respects that: the chat.disableAIFeatures setting hides chat and inline suggestions and disables the Copilot extensions, and the Chat menu in the title bar links straight to the explanation under "Learn How to Hide AI Features". Your choice survives updates.
On first launch you land in the familiar layout: Explorer, Search, Source Control, Run and Debug and Extensions in the Activity Bar on the left, editor in the middle, terminal and problems panel at the bottom. Open a folder, and VS Code recommends extensions for the languages it finds. That onboarding remains one of the friendliest in the business.
The big change: Copilot becomes the default harness
The most important change in 1.143 is not a button but a switch under the hood. The "Copilot harness" – the engine that runs AI agent sessions, built on the Copilot SDK and the Agent Host – is now the default for new sessions. In the harness picker it simply appears as Copilot, with a hover description explaining what it is. The practical upside: sessions can keep running in the background, and you can switch between the normal editor window and the separate Agents window without losing them. Version 1.142 had already started this rollout and brought back the read-only Ask agent for people who want answers without the AI touching their files.
For reviewing what an agent did, 1.143 adds a file-level working set (experimental, behind chat.editor.workingSet.enabled). A Keep action per file marks changes as reviewed and removes them from the list without touching the files on disk. More useful for everyone: the Changes view and the multi-file diff editor now offer Discard Changes not only for uncommitted changes but also for Branch Changes, Session Changes and Last Turn Changes. If an agent went off in the wrong direction, rolling back its last turn no longer means fiddling with Git by hand.
Long chats also got faster. Microsoft says chat history now loads in slices, and its own measurements (the median of three runs on macOS) show a large chat opening cold in 9.0 seconds instead of 35.6, while the time the window stays unresponsive drops from 28.3 seconds to 0.6. If you have ever watched VS Code freeze while reopening a marathon session, that is a welcome fix.
Compare Models: a bake-off inside your editor
The most fun new feature is Compare Models (Preview, enabled with sessions.chat.compareAgents.enabled in the Agents window). Pick Compare Models in the model picker, select two or more models or harnesses – Copilot, Claude and Codex are supported – along with their reasoning levels, and send one prompt to all of them. Each attempt runs as its own session, and you watch them work side by side in a grid.
Compare Models runs the same task as several parallel attempts, grouped as one comparison in the Sessions list.
An optional Judge model then produces a scorecard that also lists time and token usage, and a Synthesize button combines the best ideas – but only when you ask for it. This is a genuinely practical way to find out which model handles your codebase best instead of trusting benchmarks. The flip side is cost: four attempts use four times the AI budget.
There is also an experimental option to run a model on your own machine. A new On This Device tab in the model picker offers MAI Code 1.1 Flash, and on-device answers use no AI credits. The hardware requirement makes this a feature for workstations, though: Microsoft asks for an NVIDIA GPU with at least 96 GB of GPU memory, plus a llama.cpp build with llama-server. For typical desktop PCs, this is a glimpse of the future rather than something to switch on today.
Not only AI: diffs, diagnostics and Markdown
Developers who ignore the chat panel still get something. The multi-file diff editor can show a file list beside the diffs (experimental, multiDiffEditor.experimental.fileList), with status letters A, M, D and R and the current file highlighted as you scroll. Under More Actions > Files View & Position you choose between list or tree and left or right. Reviewing a 75-file commit suddenly feels much less like scrolling through a phone book.
The experimental file list turns large multi-file diffs into a navigable overview.
The editor also raises the limit for diagnostic highlights per file from 500 to 5,000 (each extension is capped at 1,000), and the language status indicator in the Status Bar now tells you when a limit kicks in. That matters for big generated files or strict linters, where problems previously just stopped being underlined.
If you write documentation, the Markdown editor that arrived a few versions ago is worth a look. A drop-down in the editor toolbar switches between Text Editor, Markdown Preview and the word-processor-style Markdown Editor. Version 1.142 taught it to paste images directly, rename file links across a document with F2, integrate with the Outline view and breadcrumbs, and show a copy button on code blocks. Early versions drew criticism for a jumpy cursor; in 1.142, editing mode now shows each source line on its own visual line.
How it compares
Notepad++ is the lightweight classic: tiny, instant, Windows-only and perfect for quick edits, logs and config files, but it has no integrated debugger, Git workflow or agent features. Sublime Text remains the speed champion with a beautiful, distraction-free interface, but it requires a paid licence for continued use and its extension ecosystem is much smaller. Cursor, itself built on VS Code's open-source core, goes all-in on AI from the start; with Compare Models, canvases and background agents, VS Code is clearly closing that gap while staying free at the base. Lighter editors like PSPad are still fine for occasional HTML or script editing.
Pros and cons
Pros
- Free for private and commercial use, with a huge extension ecosystem
- Compare Models makes it easy to test which AI model suits your code
- Discard Changes for Branch, Session and Last Turn Changes simplifies undoing agent work
- Faster loading of long chat sessions and higher diagnostic limits
- AI features can be hidden completely with one setting
Cons
- Release notes are dominated by AI; classic editor improvements are small
- Many new features are Preview or Experimental and need settings to enable
- On-device models require workstation-class GPUs
- Full AI use depends on a GitHub account and a Copilot plan
Our verdict
VS Code 1.143 is a solid update for anyone already using the editor's AI features: Copilot as the default harness, background sessions, Compare Models and better rollback tools make agent work more controllable and transparent. If you mostly write code yourself, the update is smaller but still worth installing for the diff file list, higher diagnostic limits and the usual stream of fixes – and the switch to hide AI entirely means VS Code does not force the new direction on you. Users who want a featherweight text editor are better served by Notepad++, and those who find the monthly AI churn exhausting may prefer Sublime Text. For everyone else, VS Code remains the best free code editor on Windows, even if its identity is shifting fast.
UpdateStar rating: 4.5/5

