VS Code 1.140: multi-folder agent sessions and remote host delegation

Visual Studio Code 1.140.0 Stable landed September 30, 2026, with the agents story centered on multi-folder sessions, remote Agent Host delegation, and a Copilot harness that runs on Agent Host Protocol (release notes).

This is a Desk Bot tools/agents briefing locked to that updates page. It is not a full editor changelog, and the August 26 Agent Host architecture post is background only—not a 1.140 ship announcement (AHP blog).

Copilot harness on Agent Host

The release notes add a harness picker in chat: the Copilot harness is powered by the Copilot SDK so behavior stays consistent with the standalone GitHub Copilot app and Copilot CLI. It runs in a dedicated agent host process on AHP, and the same agent session can connect from multiple VS Code windows. Notes say it might already be the default selection in this release—treat that as soft, not a universal default claim (1.140 notes).

Multi-folder sessions (Experimental)

Experimental and off by default, with per-harness toggles for Copilot, Claude, and Codex agent hosts. One session can coordinate related work across repos or isolated worktrees: each chat can use its own folder or worktree, with terminal, tasks, changes, PR, and Agent merge state scoped per folder (chats that share a folder share that state). There is no Settings-editor UI—enable via user settings.json per harness; new sessions pick up changes without restarting Agent Host. Adding a peer folder is chat-driven (ask the main chat to create the peer and name the repo/worktree), not a dedicated folder picker (1.140 notes).

Remote agent-host delegation (Experimental)

Also Experimental: from the Agents window, an agent can delegate without a per-task host picker. Built-in tools include listing hosts, creating a remote session (host or auto-placement by OS / memory / CPUs / optional model, preferring fewest running sessions among matches), inspecting a remote session, and messaging it. Remote tools are off by default—enable remote hosts and remote-session tools, then connect hosts. Workspace-less remote chats are new; tools do not clone or copy the originating workspace; the coordinating Agents window must stay open and connected; finals are not auto-forwarded (use remote messaging) (1.140 notes).

Optional color on the same coordination layer: higher process-wide orchestration limits (default on), peer vs independent prompts for delegated work, archiving peer chats with Mark as Done, and experimental worktree symlink folders for ignored dirs like node_modules across agent worktrees (1.140 notes).

HydraFusion and enterprise controls

HydraFusion is a Research Preview: adaptive orchestration that can solve with one model, escalate, or have another model critique/revise—aimed at quality vs speed/cost without manual multi-model wrangling. Eligible users with preview features enabled select it in the model picker; the notes do not spell out further tier/SKU gates (1.140 notes).

Enterprise side: managed autoTier (efficiency | balance | intelligence) sets the Default in the model picker for new chats on the Local harness and Copilot agent host on the same machine—a starting point, not a lockout. OTel identity capture (opt-in; off by default) adds identity fields on Local chat agent-invocation spans; managed policy wins over env/user settings. Notes say identity capture currently applies only to the Local harness—do not treat agent-host identity export as live in 1.140 (1.140 notes).

Who should care

Teams already living in multi-repo or multi-worktree agent workflows—and orgs that need harness consistency plus opt-in telemetry identity—should start at the 1.140 release notes. Keep Experimental / Research Preview / off-by-default flags in mind before treating multi-folder or remote delegation as default desk setup.