Microsoft shipped Visual Studio Code 1.129 on July 15, 2026, and followed it up with the 1.129.1 servicing update on July 17. The release is notable for two changes that move the editor in very different directions: a visual refresh hidden behind an experimental flag, and a new agent-host architecture that changes how AI coding assistants run inside VS Code.

A New Look, Test-Driven

The Modern UI preview rewrites the workbench visual language. Explorer, editor groups, the panel, Activity Bar, menus, and toolbars no longer blur into a single dark slab. Instead, each surface floats in its own container with rounded corners, distinct borders, and deliberate spacing.

Menus appear as floating panels rather than flat dropdowns. Editor groups sit inside card-like frames. Tabs and toolbar controls adopt softer edges and reduced contrast. The goal is straightforward: in a workspace crowded with split editors, a terminal, source control, and an AI chat panel, clear visual boundaries help your eyes find the right region faster.

Microsoft has been careful to call this an experiment. The feature lives behind workbench.experimental.modernUI in the stable channel. Insiders builds enable it by default, giving the team a broad test pool without forcing production users to adapt overnight. If you depend on custom themes, elaborate layouts, or accessibility tooling, you will want to test the preview on a secondary machine before switching your daily driver.

The Agent Host Arrives

While the UI tweak grabs attention, the architectural headline is the new agent host. In VS Code 1.129, supported coding-agent sessions—GitHub Copilot, Claude, Codex, and others—can now run inside a dedicated process that uses the Agent Host Protocol (AHP).

Until now, an AI conversation was tied to the editor window where it started. The agent host decouples the session from any single window. Once a session is live, multiple VS Code windows can connect to it and render its output. That means you can start a chat in one project window, clone or switch to another, and continue the conversation without losing context.

The feature is opt-in via chat.agentHost.enabled. Administrators can control it at the organization level, so many developers may not see the option until their IT team rolls it out. The host is explicitly designed to be a managed surface, and that matters: a persistent agent that can touch repositories, terminals, and source control across windows centralizes a lot of trust decisions.

Agents running through the host gain new abilities. They can list active sessions, inspect recent conversation context, create a session or chat for a subtask, and act on other sessions. Sending a message to another session still requires user confirmation, and Microsoft caps bursts of session sends to prevent a single prompt from spawning an uncontrolled chain of agent activity.

Reviewing Agent Output Gets an Editor

The separate Agents window now includes an experimental editor panel. Enable sessions.layout.singlePaneDetailPanel, reload, and any files or diffs produced by an agent open in a docked editor alongside the chat. A shared tab bar, inline and side-by-side diff views, and persistent open editors across session switches make the review process feel like normal editor work.

This is a direct response to a known friction point in AI-assisted development: generating code is fast; reviewing it thoroughly is not. By placing agent output inside an interface that behaves like VS Code’s main editor, the team is trying to reduce the gap between seeing a suggestion and actually inspecting it.

The panel does not replace code review, branch protections, or testing. For teams using Git worktrees, the new session flow exposes worktree isolation as a checkbox, making it simpler to contain an agent’s changes in a separate folder before merging.

The Terminal Gets a Bang Prefix

A small but telling addition: prefix a prompt with ! in an agent-host chat, and VS Code treats it as a terminal command rather than a natural-language instruction. The chat pane becomes a more direct control surface, useful for experienced developers who bounce between conversation and command line.

Microsoft has limited the command capability to agent-host sessions in the editor and the Agents window. It is not a general-purpose terminal replacement, and its risk profile depends on the agent harness, workspace trust, and what you type after that exclamation mark.

Other Notable Changes

  • Bring Your Own Key models: Copilot’s agent harness now supports models configured via BYOK.
  • GitHub Enterprise authentication: Copilot can authenticate through the agent host for GitHub Enterprise accounts, addressing a sign-in gap that previously blocked enterprise users.
  • Skills conversion tool: An experimental converter turns older .prompt.md files into reusable agent skills, useful for teams that have built up prompt libraries.

What This Means for Windows Developers

If you use VS Code on Windows 10 or 11, the practical impact today is minimal unless you choose to flip a few toggles. The Modern UI is a visual experiment; turn it on to help Microsoft harden it, but don’t adopt it as a team standard yet. The agent host is a more profound shift, but for now it remains an opt-in preview that many organizations will keep disabled.

Power users who work across multiple repositories and windows have the most to gain from the agent host. The ability to maintain one AI session across several editor instances can streamline complex tasks where context spans projects. But that same power demands careful configuration: review workspace trust settings, agent-host policies, and the approval model before enabling the feature on a machine you rely on for production work.

Administrators distributing VS Code should note that both workbench.experimental.modernUI and chat.agentHost.enabled can be set via policy. There is no reason to force the Modern UI onto a team’s stable configuration in its current state. For the agent host, the decision is more nuanced: if your developers use AI assistants extensively and you have a way to monitor and govern agent actions, enabling the host in a controlled pilot could surface valuable workflow improvements. If not, keep it off and let the feature mature in Insiders.

How to Enable the New Features

  1. Update to Visual Studio Code 1.129 or 1.129.1.
  2. Modern UI: Open Settings (Ctrl+,), search for modernUI, check the box for Workbench > Experimental: Modern UI, and reload the window.
  3. Agent host: Search Settings for agentHost, set Chat: Agent Host Enabled to true, and reload. Note that this option might be grayed out if your organization disables it via policy.
  4. Agent editor panel: After enabling the agent host, search for singlePaneDetailPanel, set Sessions > Layout: Single Pane Detail Panel to enabled, and reload.

What to Watch

Microsoft has a pattern of incubating significant editor changes behind workbench.experimental flags before promoting them. The Modern UI may remain experimental for several releases while the team gathers telemetry and feedback. The agent host feels like a bigger bet: it formalizes a protocol and process model that could become the default way AI sessions run in VS Code. If that happens, extensions that integrate with AI tools will need to adapt, and enterprise administrators will need clear documentation on auditing agent activity.

The next few Insiders cycles will tell us whether the agent host stays an opt-in playground or starts to leak into the default experience. For now, VS Code 1.129.1 offers a capable preview that rewards tinkering but doesn’t demand attention from users who just want to write code.