Visual Studio Code 1.143 with Copilot Default, On-Device Models, and More
Microsoft's latest stable release folds its AI-agent strategy into the defaults, nudging users toward a single Copilot-backed runtime while shipping previews for on-device models and side-by-side comparison.
Visual Studio Code 1.143 is the release where Microsoft finally stopped treating the Copilot harness as a suggestion and made it the rule. Open a new agent session in the stable build and Copilot is checked by default. There's no toggle to flip. It just happens. Those sessions run on the shared Copilot SDK and the Agent Host runtime, so they survive folder closes, sync between the editor and the dedicated Agents window, and stay alive in the background.
There's a reason this shift landed now.
The release cadence tells the story. In 2025, VS Code put out a steady 12 minor versions a year, running 1.97 through 1.108. In 2026 it sped up hard. Between 1.110 in February and 1.143 in October, Microsoft shipped roughly 33 versions, or about four a month. That near-weekly loop makes a "stable" release feel more like a snapshot of a nearly continuous update stream. Features land in Insiders first, then roll to Stable over days or a couple of weeks.
1.143 sits on top of roughly a year of agent work. Late 2025 let you keep a local agent chat running in the background. Spring 2026 added a Claude harness and a Codex harness alongside Copilot, plus a unified experience across the editor, the CLI, and the standalone app. Then came the Agent Host in July, a dedicated process that owns sessions so they outlive the window you opened them in. 1.143 defaults the Copilot harness and lets the surrounding tooling catch up.
The through-line is a single, SDK-backed runtime that behaves the same across all of Microsoft's Copilot products. Claude and Codex are still here. They're just increasingly the alternatives.
Copilot gets the default seat
When a new session starts, the harness picker shows Copilot selected. That's the headline. A smaller change sits beside it: file-level working-set support, still experimental. Feedback said approving or rejecting agent edits in the editor got fiddly under the Copilot harness, so Microsoft added a working-set bar above the chat input. Mark a file Keep and it's flagged as reviewed. The edits stay on disk, and the file reappears if the agent changes it again.
Then there's the worktree cleanup tool. Run "Chat: Open Worktree Cleanup" and you can see how much disk space inactive agent sessions are eating, filter by inactivity, and delete the ones you don't need. Active, running, pinned, and needs-input sessions are protected. That's after many of us have forgotten we've got a half-dozen worktrees parked on our drives.
The performance fixes matter too. Opening a giant Agent Host chat could previously freeze the window for up to 35 seconds while history restored. Microsoft now slices history into small chunks. Its median-of-three benchmarks (macOS, real chats) show a cold open dropping from 35.6 seconds to 9.0, with the window unresponsive for 28.3s of that. Now that's 0.6s. Microsoft framed the work as "continuing to improve performance and reduce latency," citing a fix for the lag that shows up when MCP servers don't finish loading their tools.
There are also a couple of previews worth flagging. Compare Models runs the same prompt across two or more models in parallel, mixing Copilot, Claude, and Codex with different reasoning levels, then combines the outputs. There's an optional Judge model that scores each attempt. It's arguably the most genuinely useful preview here, though it burns credits for every model it runs.
On-device models is the flashier bet. Microsoft is targeting hardware like the Surface Laptop Ultra with RTX Spark and up to 128 GB of unified memory. You download MAI Code 1.1 Flash from a new "On This Device" tab and run it locally. Those responses don't use AI credits, and that's the hook. The catch is the bar: you need an NVIDIA GPU with 96 GB or more of memory. Firmly enthusiast territory.
Microsoft pitches MAI Code 1.1 Flash as having "outperformed Claude Haiku 4.5 and GPT-5.4 Mini" on quality while using fewer tokens. Keep in mind those numbers are Microsoft's own aggregated usage data, not independent benchmarks.
The enterprise play is the real story
A lot of 1.143's substance points at IT admins. Much of it lives in managed settings, not user preferences.
New chats now honor the organization's managed model defaults even when a user has set their own. There's a new dynamic header provider for OpenTelemetry exports, so admins can push short-lived auth tokens without developers hand-editing anything. Managed plugins install automatically before a session runs, taking precedence over trusted-repository settings and user choices.
Then comes the enforcement side. When policy requires the Copilot harness, Local stays visible but disabled, with a "Learn why" link. Existing Claude and Codex sessions you already started can keep running, but you can't create new ones. The Local controls are explicitly temporary, tied to Local's retirement, which tells you the consolidation isn't finished yet.
If you want to try the experimental features, most sit behind setting toggles like chat.editor.workingSet.enabled and chat.copilot.localModel.enabled. Preview features are opt-in and can change. Head here to the release notes for the full settings cheat sheet.
One caveat before you upgrade. Several of the performance and quality claims come from Microsoft's own aggregated data. And the on-device feature is hardware-gated and experimental. Treat today's defaults as reversible, and watch Insiders.
