Microsoft shipped Visual Studio Code 1.143. The Copilot harness, built on the Copilot SDK, is now the default for new sessions and works across the editor and Agents window. A Compare Models preview runs several models on one task with an optional Judge model, an experimental mode runs MAI Code 1.1 Flash on a local NVIDIA GPU, and large chats open much faster.
Key Takeaways
- ✓Default harness: new sessions use the Copilot harness built on the Copilot SDK, which can run in the background and move between the editor and the Agents window
- ✓Compare Models (Preview): one prompt runs 2 or more models in parallel, mixing Copilot, Claude and Codex harnesses, with an optional Judge model scoring time and token use and a Synthesize step
- ✓On-device models (Experimental): download MAI Code 1.1 Flash to run locally, labeled On-device with no AI credits used; needs an NVIDIA GPU with 96 GB+ memory
- ✓Performance (vendor figures): cold-opening a ~1,150-request chat dropped from 35.6 s to 9.0 s, and window freeze time from 28.3 s to 0.6 s
- ✓Also: the editor diagnostic highlight limit rises from 500 to 5,000 per file, and the Windows installer can install GitHub Copilot CLI
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
VS Code has carried several agent harnesses (Local, Copilot, Claude, Codex) with uneven behavior, and users have reported slow-opening long chats and agent worktrees eating disk space. On Oct 21 the official VS Code account announced release 1.143 (release notes dated Oct 19), which unifies the agent foundation and adds multi-model tooling.
How it works
New sessions default to the Copilot harness, built on the Copilot SDK that Microsoft shares across Copilot products. Sessions run through the Agent Host, keep working in the background, and move between the editor and the Agents window. A Compare Models preview in the Agents window runs several models on one prompt, each in its own session and worktree, mixing Copilot, Claude and Codex harnesses and reasoning levels, with an optional Judge model and a Synthesize step. An experimental on-device mode downloads MAI Code 1.1 Flash and runs it through llama.cpp's llama-server; on-device replies are labeled and use no AI credits. When enterprise policy requires it, the Local harness is disabled, and Microsoft says Local is being retired.
Benchmarks
No model benchmarks were published. Microsoft did publish chat-opening measurements (vendor figures, median of three runs on macOS): a ~1,150-request chat with ~530 MB of history now cold-opens in 9.0 s instead of 35.6 s, with the window unresponsive for 0.6 s instead of 28.3 s, and re-opens in 0.6 s instead of 26.7 s. The per-file diagnostic highlight limit rises from 500 to 5,000.
Getting started
After updating, new sessions use the Copilot harness automatically. To compare models, enable sessions.chat.compareAgents.enabled in the Agents window and pick Compare Models in the model picker. On-device models need GitHub sign-in and an NVIDIA GPU with 96 GB or more memory; install a compatible llama.cpp build, set chat.copilot.localModel.llamaServerPath, and enable chat.copilot.localModel.enabled. Run Chat: Open Worktree Cleanup to remove inactive agent worktrees.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.