VS Code 1.143: Copilot Harness Default and Compare Models

VS Code 1.143 Makes the Copilot Harness the Default for New Agent Sessions

Microsoft released Visual Studio Code 1.143 in the third week of October 2026, followed by the 1.143.1 patch on 23 October. The most important change is architectural: new agent sessions now use the Copilot harness by default. The Copilot harness is built on the Copilot SDK and runs through the Agent Host, giving GitHub Copilot's agent capabilities a shared foundation across Copilot products. Sessions can keep running in the background and can be opened from both the editor window and the separate Agents window. The release also adds a way to run several AI models against the same prompt, an experimental on-device model that does not consume AI credits, much faster loading for large chat histories, and new enterprise controls.

Compare Models: Run Two or More AI Models on One Prompt

Compare Models is a preview feature in the Agents window, enabled with the sessions.chat.compareAgents.enabled setting. It lets a developer send one prompt to two or more models or harnesses at the same time and review the results side by side. An optional Judge model can evaluate the competing answers, and a Synthesize action combines the results into one response. For teams that are still deciding which model to standardise on, this turns model evaluation into an everyday editor workflow rather than a separate benchmarking exercise.

More Control Over What the Agent Changes

VS Code 1.143 gives developers finer control over agent edits. An experimental file-level working set, enabled with chat.editor.workingSet.enabled, adds a Keep action to files. Kept edits stay on disk, and a file reappears in the working set if the agent changes it again. The Discard action, previously limited to uncommitted changes, now also works on Branch Changes, Session Changes and Last Turn Changes. Together, these make it easier to accept part of an agent's work and roll back the rest without reaching for Git commands.

An experimental Worktree Cleanup command lists inactive agent worktrees so they can be reviewed and removed. Active, running, pinned and needs-input sessions are protected from cleanup. For developers who run many parallel agent sessions, this addresses the disk space that old worktrees quietly consume.

Large Chats Open Four Times Faster

Microsoft measured a cold open of a large chat with about 1,150 requests and roughly 530 MB of history. Time to open fell from 35.6 seconds to 9.0 seconds, and the period in which the editor was unresponsive fell from 28.3 seconds to 0.6 seconds, based on the median of three macOS runs. Long-running agent conversations were one of the most common sources of editor sluggishness, so this is a practical quality-of-life improvement.

On-Device AI Without Credits and BYOK Voice Mode

An experimental On This Device tab lets developers download MAI Code 1.1 Flash and run it locally. On-device responses do not use AI credits. The requirements are steep: a llama.cpp build, GitHub sign-in, and an NVIDIA GPU with 96 GB or more of GPU memory, so this is aimed at AI workstations rather than ordinary laptops. A separate preview adds bring-your-own-key Voice Mode, letting developers use their own OpenAI key for voice interaction with agents.

Sandboxing also improves. A new Sandbox: Manage Trusted CA command installs, rotates or removes trust for the sandbox certificate authority on Windows and macOS agent hosts. Where managed policy allows it, developers can turn off sandboxing for terminal commands in a single Copilot harness session.

Enterprise Controls and Editor Improvements

For organisations, new Copilot harness chats now use organisation-defined default models even if an individual has set a different default. When Copilot is required by policy, the Local harness stays visible but disabled in the picker. Required agent plugins install or update automatically, and a telemetry headers provider setting refreshes OpenTelemetry headers for Copilot Agent Host sessions. Outside AI, the per-file diagnostic highlight limit rose from 500 to 5,000, and an experimental file list for the multi-diff editor shows added, modified, deleted and renamed files as a list or tree. The installers now offer to install the GitHub Copilot CLI on Windows, Linux and macOS.

What VS Code 1.143 Means for Indian Development Teams

For Indian software companies and IT services firms, VS Code 1.143 strengthens the case for treating AI agents as governed, team-wide infrastructure rather than personal tools. Managed default models and required plugins let engineering leads standardise which models every developer uses on a client project, which simplifies both cost control and client data agreements. Compare Models gives delivery teams a quick, evidence-based way to choose the right model for each codebase. Granular discard and file-level working sets make it safer to let agents work on production code, because reviewers can keep exactly the changes they approve.

The Bottom Line

Visual Studio Code 1.143, released in October 2026 with a 1.143.1 patch on 23 October, makes the Copilot harness the default for new agent sessions. It adds a Compare Models preview with optional Judge and Synthesize steps, a file-level working set, discard support for branch, session and last-turn changes, and worktree cleanup. Large chats now open in 9.0 seconds instead of 35.6 seconds. An experimental on-device MAI Code 1.1 Flash model runs without AI credits on GPUs with at least 96 GB of memory, and new enterprise policies enforce organisation default models and required plugins. Teams should update, enable the previews on a pilot project, and set organisation-wide model policies before rolling agents out more widely.