VS Code 1.143: Copilot Harness Goes Default, Plus Model Comparisons

VS Code 1.143 is out, and the headline change is a routing one: new agent sessions now default to the Copilot harness instead of VS Code's built-in Local harness. The October update, released October 22, 2026, also adds a preview "Compare models" mode that runs one coding task across Copilot, Claude and Codex at once, plus experimental tools for reclaiming disk space from old agent sessions and running a small AI model entirely on your own GPU.

What's new in VS Code 1.143

VS Code ships a dot release roughly every month, and 1.143 is the October 2026 build. Microsoft published the release on GitHub on October 22, 2026, with a same-day 1.143.1 patch, and the full change list sits in the official 1.143 release notes. Most of the update is about AI agent sessions inside the editor: which model runs them by default, how to compare several models on the same task, and how to manage the disk space and performance costs that come with running a lot of agent sessions over time. None of it requires a fresh install — it lands through VS Code's normal auto-update pipeline on Stable.

Copilot harness is now the default

The biggest change in 1.143 is under the hood. VS Code's chat and agent features can run on different "harnesses" — the execution layer that turns a prompt into file edits, terminal commands and test runs. Starting with this release, new sessions default to the Copilot harness, which Microsoft has rebuilt on its Copilot SDK so that behavior stays consistent across other Copilot products rather than diverging inside VS Code specifically.

Sessions launched this way run through what Microsoft calls the Agent Host, which means a session can keep working in the background and be picked back up from either the main editor or the dedicated Agents window. Anyone who had pinned a different default harness, including the Local harness that enterprises can disable by policy, keeps that choice — the change only affects what a brand-new session starts with.

VS Code's own agents documentation describes a harness as the layer that decides how a model's suggestions turn into actual edits, terminal commands and test runs — swapping the default doesn't change which model you're chatting with, just the plumbing underneath it. If GitHub Copilot isn't already installed, it ships as a Marketplace extension that most VS Code installs already bundle by default.

Compare models: run one task across several models

The second major addition is a preview feature called Compare models, available in the Agents window once the sessions.chat.compareAgents.enabled setting is on. Instead of picking one model and hoping it gets a task right, you select "Compare Models" in the model picker and choose two or more entries — which can mix Copilot, Claude and Codex harnesses at different reasoning levels — then send a single prompt that kicks off every attempt in parallel, each in its own session with its own worktree and its own file changes.

An optional Judge model can then score the attempts on a scorecard that reports time and token usage, its reasoning, and validation evidence, with links back to each individual attempt. From there, a "Synthesize" step — run only if you ask for it — builds on whichever attempt the Judge recommended, pulling in ideas from the runs that didn't win. Microsoft positions this as a way to spend more compute on a task you only want to run once, rather than a replacement for picking a single default model day to day.

Because every attempt gets its own worktree and its own set of file changes, nothing about Compare models touches your actual working directory until you decide which result to keep — you're reviewing several completed attempts side by side rather than watching several models argue over the same files in real time. That isolation is also why the feature pairs with the worktree cleanup tool described below: running comparisons regularly means accumulating several throwaway worktrees per task, and those need somewhere to go once you've picked a winner.

Reclaiming worktree storage

Every agent session that touches files gets its own Git worktree, and those can pile up unnoticed. The new experimental "Chat: Open Worktree Cleanup" command, run from the Command Palette, lists inactive agent-session worktrees, lets you filter them by how long they've sat idle, shows how much space each one holds, and deletes the ones you select. Sessions you clean up are marked done rather than destroyed outright, so they can be restored later; anything still active, running, waiting on input, or pinned is left alone. The cleanup view can also offer to clear out sessions whose pull requests have already merged, and a new setting, chat.agentSessions.sessionStorageCleanupSuggestion.enabled, controls whether VS Code proactively suggests a cleanup once storage use grows.

Running a model on your own GPU

For developers who want agent features without sending prompts to the cloud, 1.143 adds experimental support for on-device models. The first one available is MAI Code 1.1 Flash, which you download from a new "On This Device" tab in the model picker and then use inside Copilot harness sessions. Responses from the local model are labeled "On-device" in the UI and don't draw against your Copilot AI credits; an "Use in Auto" option can route suitable requests to the local model automatically, with cloud fallback when it can't handle a request.

This isn't something most laptops can turn on today. According to Microsoft's own release notes, it requires signing in with GitHub, installing a llama.cpp build, pointing VS Code at the llama-server executable via chat.copilot.localModel.llamaServerPath, and — the real gatekeeping requirement — an NVIDIA GPU with at least 96 GB of GPU memory. Microsoft cites Microsoft's own Surface Laptop Ultra, with up to 128 GB of unified memory, as a device that qualifies, which gives a sense of the hardware tier this feature currently targets.

Faster large chats

If you've kept long-running agent conversations open for weeks, 1.143 should feel noticeably snappier. VS Code now restores chat history in small slices instead of one blocking operation, and Microsoft's own benchmarks — run three times on macOS against real development chats — show the difference:

Those numbers are Microsoft's own measurements from its release notes, not independent benchmarks, but the direction is consistent across every scenario the company published: cold opens, reopening a chat you switched away from, and reloading the whole window.

Other changes worth knowing

A handful of smaller changes round out the release. An experimental "Working set" mode lets Copilot harness editor sessions mark files as Keep, taking them out of the active working set while leaving the edits on disk. Discarding changes now works on Branch, Session and Last Turn changesets, not only Uncommitted Changes. Detached background shells stream their output live and can be stopped from the UI. An experimental vertical Session Activity Bar groups Changes, Files, Browser and canvas views, though it stays off by default in the Stable release. The per-file diagnostic highlight limit rose from 500 to 5,000, which matters on files with heavy linting. On the setup side, the Windows installer and the macOS PATH prompt can now install the GitHub Copilot CLI alongside the editor, and a preview Voice Mode supports bring-your-own OpenAI live models.

Developers who rely on Model Context Protocol (MCP) servers for custom agent tools aren't affected by the harness change directly — MCP tools still plug into whichever harness is active — but it's worth checking any session-startup automation against the new default before rolling 1.143 out to a team.

Local, Copilot, or something else?

VS Code has supported more than one harness for a while, but 1.143 is the first release where the choice is made for you by default rather than left as whatever you happened to pick during setup. The Local harness runs entirely inside VS Code's own agent loop without routing through Copilot's backend, and some enterprises keep it as the only option available to staff through policy — the release notes explicitly call out a "Local harness disabled by policy" setting for IT administrators who want to lock that down rather than let users switch to Copilot or a third-party harness. Individual developers who prefer Local, or who already have Claude or Codex set as their default through a different extension, aren't forced to switch; the new default only applies the first time a profile starts a session with no harness preference saved.

The practical effect for most people is that anyone who installs VS Code fresh, or who never touched the harness setting, now gets Copilot's SDK-backed execution layer out of the box — including the background Agent Host behavior — rather than whatever the previous default happened to be.

How to update to VS Code 1.143

VS Code auto-updates by default, so most users will already be on 1.143 or the 1.143.1 patch. To check manually, open the Command Palette and run "Check for Updates," or look at Help > About for the version number. Anyone on a managed or enterprise build should confirm with their IT team, since 1.143 also ships enterprise controls for managed model defaults, a policy-disabled Local harness, managed extension auto-install, and dynamic OpenTelemetry headers that administrators may want to configure before the new Copilot-harness default reaches their users. Windows users mid-upgrade cycle may also want to revisit the latest Windows 11 update for any interaction with installer or PATH changes.

What's next

Compare models, on-device models, the multi-diff file list and worktree cleanup are all flagged experimental or preview, which means Microsoft can change or drop them before they reach Stable by default. The clearest signal to watch for is whether Compare models graduates out of preview in the next monthly release, since that's the feature most likely to change how developers actually choose a model day to day. VS Code's release notes and the project's GitHub releases page are the two places Microsoft confirms that kind of change first.

Written byFelix Moreau

Felix Moreau writes Pandromeda's software coverage and how-to guides. He covers Windows, macOS and Linux updates, the apps people rely on, emulators and developer tools, and turns official documentation into clear, numbered steps that work on the current version.

More from Felix MoreauAll Software & Guides coverage