VS Code stopped being an editor with a chat panel a while ago. Two releases in eight days, 1.146 on November 11 and 1.147 today, make it official: the product is now a control room for agents, and the features that matter are about managing attention, not typing code.
There’s also a hard deadline buried in the notes. If you still run the Local harness, you have until February 2027.
What shipped
| Release | Date | Headline features |
|---|---|---|
| 1.146 | Nov 11 | Inbox mode, chat windows, Copilot Fusion in model comparison, Skills over MCP, Local harness deprecation |
| 1.147 | Nov 18 | Send Later, drafts, model routers, browser integration, HTML previews in chat |
Most of it is labeled Experimental or Preview and sits behind settings. That’s fine. The direction is more interesting than any single toggle.
The Local harness is on a three-month clock
The Copilot harness, built on the Copilot SDK, has been the default for new sessions since 1.143. As of 1.146, the Local harness is deprecated and will be removed in February 2027. Until then it’s functionally frozen: new features target the Copilot harness and aren’t guaranteed to work on the old one.
The enterprise-managed vscode.temporaryDefaultToLocalHarness setting, introduced in 1.142 as an escape hatch, goes away too.
Two reasons I think this is the right call. First, Copilot harness sessions run through the Agent Host, so they keep going in the background and show up in both the editor window and the Agents window. Second, for organizations, only the Copilot harness enforces managed sandbox and permission policies. If you’re a platform team, “we can’t enforce policy on the old path” is a deadline you set yourself. Audit who’s still on Local now, not in January.
Attention management is the new feature category
I’ve argued in my agentic development maturity curve piece that the bottleneck moves from writing code to supervising it. These two releases are VS Code admitting the same thing.
Inbox mode
Enable sessions.list.inboxEnabled and the Agents window sessions list groups work by what needs you: sessions needing attention on top, then Working, Ready to merge, On hold, and Done. In 1.147 you can also mark a session as acknowledged. It leaves the Inbox, but it isn’t archived or marked done, and it returns the moment there’s new activity. Workspace and environment pickers at the top filter every section.
That acknowledge-but-don’t-close state is a small detail that tells me the team is watching how people actually run five agents at once.
Chat windows
With sessions.chatWindows.enabled (on by default in Insiders, off in Stable), a chat can live in its own window. Right-click a session and choose Open in > New Window, or drag it out. A window can hold several chats, with their editors and terminals arranged around them, and it reopens after a restart. Canvases follow their chat into the new window.
Send Later and drafts
This is my favorite 1.147 feature. With sessions.sendLater.enabled you can schedule a chat request for a specific time. You can also just say it in plain language:
Check the build in 10 minutes.
The agent schedules a one-time, text-only request in the current chat. If the agent is busy at send time, the request waits its turn instead of interrupting. The caveat is real: it only works for sessions on This Computer, requires automations to be enabled, and the local agent host has to be running. Scheduling doesn’t wake a sleeping computer, though a missed request is processed when the host comes back.
Drafts (sessions.drafts.enabled) complement it. Instead of one unsent request in the new-session box, you keep several, each remembering its workspace, agent, model, attachments, and inline references. A draft is only removed after the send is accepted, so a failed send doesn’t eat your prompt.
Together they make asynchronous agent work feel like email: compose now, deliver later, triage in an inbox. I don’t think that’s an accident.
Model routers: stop picking models by hand
1.147 adds experimental model routers. You write a YAML file in ~/.agents/model-routers/ that maps kinds of work to models, and a decision model, gpt-6-luna-decisions, picks the route per request:
version: 1
description: Opus to build, GPT-6 Luna for quick questions.
routes:
- model: opus
when: Building or changing code over many steps, such as a feature, a bug fix, or a refactor.
- model: gpt-6-luna
when: Quick answers and small chores, such as where something is or what a setting does.
Turn it on with "chat.copilot.modelRouter.enabled": true in your user settings.json, or run /model-router in a Copilot harness chat and describe the routing in prose. The router then shows up in the model picker.
My take: this is cost and latency governance disguised as a convenience feature. Most of us waste premium-model requests on “where is this setting?” questions. Routing by task type is the same instinct behind the context discipline I wrote about in context engineering: give the model what the job needs, no more. It’s early, and Microsoft says so, but I’d start writing routes now.
Model comparison grows up
1.146 let you add Copilot Fusion as an attempt, judge, or synthesizer in Compare Models, and let judges and synthesizers run without a workspace. With sessions.chat.compareAgents.tools.enabled, an agent can discover models, launch comparisons, and read results itself. In 1.147, comparisons you start through an agent appear under that chat, and hovering shows the original repo, isolation, base branch, models, effort levels, judge, and synthesizer. If you’ve ever tried to reproduce a comparison from memory, you’ll appreciate that.
Security and extensibility changes worth noticing
- Granular permission prompts (1.146). The Copilot harness used to offer “Allow in this Session” for a web fetch, which meant any URL. Prompts now default to specific domains, commands, and directories. A quiet but important fix.
- Masked environment variables in the sandbox (1.146). Sandboxed tools receive placeholders; real values go only to approved HTTPS destinations. That’s the right model for secrets and agents.
- Customizable assisted permissions (1.147). You can tell the permission reviewer which actions you consider safe and which need approval.
- Skills over MCP (1.146 preview). With
chat.agentHost.mcpSkills.enabled, MCP servers can supply skills to the model automatically, after you approve them on first connect. They appear on the Skills tab of Agent Customizations. - Extensions in the Agents window (1.146). You can install and manage extensions without leaving it, with warnings when an extension contributes views or debuggers the window doesn’t support.
Smaller things I’ll actually use
/codein Copilot CLI. Run it and the CLI session, with history, model, reasoning effort, and mode, opens in VS Code. It needs CLI 1.0.107+ with experimental features on, and CLI-only approvals don’t carry over. I covered the CLI’s pace in Copilot CLI’s biggest week yet; this closes the loop between terminal and editor.- Dev Container state volumes. Chat sessions survive container rebuilds if you enable
dev.containers.stateVolume.enabledand rebuild once. - Browser integration. In 1.147 preview, agents can drive your default browser. Per AI/TLDR’s summary, it needs the GitHub Copilot browser extension, a local Copilot harness session, and Node.js 22+.
- Inline HTML previews and charts in agent responses, plus Mermaid 12 rendering in chat, previews, and notebooks.
- Cron schedules for local automations arrived in 1.145, which is what Send Later builds on.
For broader coverage, Visual Studio Magazine’s 1.146 write-up is a decent second read.
What it means
Everything here assumes you’re running many agents, in parallel, over hours. The primitives are the ones a human manager needs: an inbox, scheduling, drafts, delegation rules (routers), and permission boundaries. None of them is about code completion.
That has a practical consequence. Your VS Code settings are becoming operational config. Model routers, permission reviewers, and sandbox masking are the kind of thing a team should standardize, version, and review, exactly like CI config.
Bottom line
Migrate off the Local harness before February. Try Inbox mode and Send Later this week, and write one model router for your most common split of hard versus trivial work. VS Code’s direction is clear: the editor is the cockpit, and the agents are the crew. Expect the next releases to be about coordination between those agents, not about any one of them.