VS Code 1.140 lets one agent session manage multiple repos

Microsoft released Visual Studio Code 1.140 on September 30. Nearly every change in this release is aimed at agents: the architecture under Copilot's agent runtime has been rebuilt, a single session can now work across multiple folders at once, agents can hand off tasks to a remote machine, and a new multi-model orchestration research preview called HydraFusion has shown up.

The runtime moves into its own process

VS Code 1.140 rebuilds Copilot's agent runtime (what Microsoft calls the Copilot harness) around the Agent Host Protocol (AHP), running it in a separate process. The immediate payoff: multiple VS Code windows can now connect to the same agent session, and closing one window no longer kills the session with it.

Built on top of that runtime, multi-root sessions ship as an experimental feature. A single session can now work across several repositories or isolated worktrees at once, with each chat getting its own folder, terminal, tasks, and change history. Three settings turn it on:

  • chat.agentHost.copilotAgent.multiRootEnabled
  • chat.agentHost.claudeAgent.multiRootEnabled
  • chat.agentHost.codexAgent.multiRootEnabled

Judging by the naming, VS Code's agent host gives Claude and Codex agents the same toggle it gives its own Copilot agent.

Delegating to remote machines, with automatic model switching

Remote delegation is also experimental, controlled by chat.remoteAgentHostsEnabled and chat.remoteSessions.tools.enabled. Turning it on gives the agent four new tools: list available remote hosts, create a remote session, check a session's status, and send a message to a remote session. In practice, the local agent acts as a dispatcher, offloading time-consuming work to a remote machine.

HydraFusion is an adaptive system that decides which model to use and which workflow to run: a simple task gets handled by one model end to end; a harder one gets escalated to a stronger model; a separate model can also be brought in to review and fix the result. Microsoft is labeling it a research preview and hasn't published its scheduling rules or any performance numbers.

A few smaller time-savers

  • Worktrees: a new git.worktreeSymlinkFolders setting lets multiple worktrees share git-ignored directories such as node_modules, so dependencies don't need to be reinstalled for every new worktree (experimental).
  • Dev Containers: containers used for agent sessions now stop automatically after 5 minutes of idle time, adjustable via chat.agentHost.devContainer.idleTimeout; VS Code Server and CLI downloads are now cached and reused.
  • Enterprise controls: admins can use autoTier to set a default tier for the Auto model — efficiency, balance, or intelligence — while developers can still override it themselves.
  • MCP configuration: supported both as a global config at $COPILOT_HOME/mcp-config.json and as a workspace-level .mcp.json.

What it means for dev teams in China

Many teams in China split their codebase into multiple repositories by service — one each for frontend, backend, and shared components. Getting an agent in VS Code to change a feature across repositories used to mean either juggling multiple windows or temporarily stitching several repos into one workspace. Multi-root sessions are built around exactly this pattern, which should make cross-repo changes — updating an interface on both frontend and backend at once, for example — noticeably smoother.

Sharing node_modules across worktrees helps even more for developers on networks that are sensitive to this kind of friction. Most dependency installs in China go through mirror sources, and reinstalling for a mid-sized frontend project can easily take several minutes each time. Running three or four worktrees in parallel so an agent can try out different approaches used to mean installing dependencies three or four times over; now it only needs to happen once.

autoTier on the enterprise side is mainly about cost control. Locked to efficiency by default, it routes everyday completions and Q&A to cheaper models, with a manual switch to intelligence when needed. The release notes don't say which models are actually available to accounts in China, or whether HydraFusion will ever route to models beyond Claude or Codex — that depends on what's actually selectable in each account.

Most of these features carry an "experimental" label and are off by default. Turning them on requires a manual flip in settings, and it's worth running them on one or two repositories for a while before rolling them out across a whole team.

Sources: Visual Studio Code 1.140 release notes, CocoLoop; setting names and defaults follow the official release notes.