VS Code 1.144: Copilot Fusion GA, Computer Use, Agents

Microsoft released VS Code 1.144 on October 28, making Copilot Fusion generally available across all paid Copilot plans, shipping experimental computer use behind the `/computer on` command, and adding cloud agent sessions that persist on GitHub's compute infrastructure after a local machine shuts down. Fusion routes each request to Standard, High, or Max reasoning intensity automatically and bills every model it invokes at that model's standard rate, with GitHub running a 30% Fusion discount on paid plans through November 30, 2026; cloud sandbox compute billing starts in November 2026. The release also adds a Working Set review layer with inline Keep and Undo buttons, enabled by default via `chat.editor.workingSet.enabled`, though terminal commands are not tracked in it.

VS Code 1.144 landed on October 28 with four changes that actually matter: Copilot Fusion graduates from preview to GA, experimental computer use ships inside the main editor, image generation arrives in Copilot Chat, and agent sessions now survive closing your laptop. This is not a maintenance release. This is the version where VS Code stops being an editor with AI bolt-ons and starts behaving like a platform for coordinating agents. Copilot Fusion Is GA — and Your Bill May Surprise You Copilot Fusion is now generally available across all paid Copilot plans. It automatically routes each request to the right model at the right reasoning intensity — Standard, High, or Max — without you manually switching models. In practice, it selects the cheapest capable model for simple completions and escalates to heavier reasoning for complex refactors. There is a billing detail worth understanding before you enable this in production: Fusion bills each model it invokes at its standard rate . A single complex request can hit multiple models in one pass. GitHub is running a 30% discount on Fusion for paid plans through November 30, 2026, which makes now the right time to test it under load and establish a real cost baseline before the discount expires. One known rough edge: Fusion can ignore explicit sequencing instructions. Users have reported that telling it to “summarize the plan before making changes” gets bypassed, with edits applied immediately. If your workflow relies on a human-in-the-loop approval step before file changes land, test this thoroughly https://github.com/microsoft/vscode/issues/345010 before depending on it for production code. Agents That Run When Your Machine Is Off Cloud sessions are the structural shift in this release. You can start an agent task from GitHub Mobile at standup, steer it from VS Code at your desk, and check in from github.com at lunch — the session persists across all of them, running on GitHub’s compute infrastructure in the background. The mechanism is snapshots: when a cloud sandbox shuts down, it saves its state. Resume from any surface without losing progress. For long-running refactors or batch tasks that take hours, this changes the calculus entirely — you no longer need to keep a terminal window open overnight. Requirements: a paid Copilot plan and an organizational policy that permits cloud execution. Billing for cloud sandbox compute starts in November 2026 https://github.blog/changelog/2026-10-28-work-from-anywhere-with-a-new-connected-copilot-cloud-experience/ , so there is a brief window to evaluate actual usage before costs appear on your bill. Organizations can enforce which repositories permit cloud execution through Copilot policy settings. Computer Use: One Command Away Type /computer on in the Copilot harness to let agents interact with desktop applications — not just the editor and terminal, but actual desktop apps on macOS or Windows. Turn it off with /computer off . This is experimental, and the word “experimental” is doing real work here. Enabling computer use gives agents the ability to click through UIs, read screens, and interact with applications that have no API. The practical use case today is UI testing — ask an agent to open a browser, navigate to a page, and verify that a button renders correctly. The security surface is significant. Enable it with clear intent, not curiosity. Working Set: A Review Layer for Agent Edits Agents making file changes now surface them through a Working Set — inline Keep and Undo buttons for each edit, visible in the panel above the chat input. The setting chat.editor.workingSet.enabled is on by default in VS Code 1.144. One gap worth noting: terminal commands are not tracked in the Working Set. If your agent runs shell commands to modify state, those changes sit outside this ledger. Review terminal output separately. This is the right feature in the right direction; it just needs broader scope. Extension Publishers: December 1 Deadline Buried in the release notes but relevant to every VS Code extension publisher: Azure DevOps PATs retire on December 1, 2026 . Migrate to vsce publish --oidc from GitHub Actions — no PAT storage required. You need @vscode/vsce 4.0.2 or later. If you are still using a global PAT in your publish pipeline, the clock is running. The Bigger Picture Each of these features is useful on its own. Together, they describe a deliberate direction: VS Code is becoming a management layer for agents that operate continuously, across devices, and across tools beyond the editor itself. The Agent Host Protocol v1.0, published with this release https://github.blog/changelog/2026-10-28-new-tools-triggers-and-cost-controls-for-copilot-automations/ , lets third parties build into this model with client libraries in six languages. This is not a feature race with Cursor. It is a different architecture — one where the editor is the control plane for a fleet of agents, not the place where all the work happens. Full release notes are available at code.visualstudio.com/updates/v1 144 https://code.visualstudio.com/updates/v1 144 .