Source · release
VS Code 1.143: compare model results and the Copilot harness by default
Summary
Visual Studio Code 1.143 (with patch 1.143.1) was released on 2026-10-22. The release
notes give no date; it comes from the microsoft/vscode release feed. It makes the
Copilot harness the default for new agent sessions and adds a preview that runs one task
with several models and then compares or combines the results. It also has experimental
on-device models, auto-install of plugins required by an organization, and many changes
to the Agents window.
Key points
- Compare model results (preview): in the Agents window, run the same task with
several models or harnesses at once. Each attempt gets its own session. An optional
Judge model scores the attempts and recommends one, and Synthesize combines ideas
from them. Enable it with
sessions.chat.compareAgents.enabled. - Copilot harness by default: new sessions use the Copilot harness, built on the Copilot SDK and run through the Agent Host. Sessions can keep running in the background and be opened from both the editor and the Agents window.
- On-device model (experimental): MAI Code 1.1 Flash can run locally for Copilot
harness sessions, and these runs use no AI credits. It needs GitHub sign-in, a
compatible llama.cpp build, and an NVIDIA GPU with at least 96 GB of memory. Set
chat.copilot.localModel.llamaServerPathandchat.copilot.localModel.enabled. - Managed plugins: administrators can require agent plugins with
enabledPluginsin managed settings (for example"my-plugin@my-marketplace": true). Sessions then install or update them automatically, for that session only. - Session tools: discard changes per branch, session, or last turn. An experimental command removes inactive agent worktrees to free disk space. Detached background shells get output and stop controls. Sandboxing can be toggled per session where policy allows it. Long chats open much faster.
- MCP: chat no longer waits as long for MCP servers that never finish loading their tools.
- Install: the GitHub Copilot CLI can now be installed through the VS Code installer.
Why it matters
Running several models on the same task and having a judge pick or merge the results is now built into a mainstream editor. Before, people did this by hand with worktrees. Organizations can now push required agent plugins to every developer.