Visual Studio Code 1.147 hit Stable on November 18, and it is almost entirely an agent release. If you use the Agents window, this one changes how you schedule work, pick models, and even where your agent is allowed to operate. One honest caveat up front, straight from the release notes: nearly everything here ships as Experimental or Preview, off by default. Microsoft is laying down the scaffolding for agent-heavy workflows faster than it is turning the features on.
Browser Use: the agent leaves the editor
The headline feature is Browser Use (Preview). It lets an agent interact directly with your default browser — Google Chrome or Microsoft Edge — instead of working only inside VS Code. The practical difference: the agent can reuse your existing authenticated sessions and interact with web pages outside the editor, complementing VS Code's integrated browser.
It is not a one-click feature. It requires the GitHub Copilot browser extension, a Copilot harness session running on your own machine, and Node.js 22 or later on your PATH. That setup friction is worth noting, because it tells you who this is for right now: people already running local agent sessions, not casual Copilot Chat users.
Model routers: describe the split once, stop picking per request
Model routers (Experimental) replace per-request model picking with a YAML file in ~/.agents/model-routers/. Each route pairs a model with a plain-language description of the work it should get, and a decision model — gpt-6-luna-decisions — matches each incoming request to a route. The release notes' own example sends building and refactoring work to Opus and quick questions to GPT-6 Luna. You can author a router with the /model-router command in a Copilot harness chat.
This is the cost-versus-quality tradeoff made declarative. The catch, also visible in the notes: routing is chosen by a model, so a misrouted request costs you either money or quality — and routers are Copilot-harness-only for now.
Send Later and drafts: the agent gets a schedule
Send Later (Experimental) lets you postpone a chat request until a specific time, from a menu on the send button or by asking the agent in plain language — "Check the build in 10 minutes." Drafts (Experimental) let you keep several unsent requests side by side, each with its own workspace, agent, and configuration; they survive reloads and restarts.
Two limits matter. Scheduled requests work only for sessions on This Computer and require automations to be enabled — and scheduling does not wake a sleeping computer. A missed request is processed once the host is available again. So this is a scheduler for a machine that stays on, not a cron replacement.
Visual answers and assisted permissions
Agents can now show visual responses inline in the Agents window: HTML previews via the /html-preview skill for reports and mockups, and native visuals via the /vscode-ui skill for metrics, badges, and bar charts rendered in your theme. Both are off by default.
Assisted permissions get a useful upgrade: chat.permissions.assisted.instructions lets you tell the LLM reviewer which actions you consider safe and which need approval ("Always ask before installing packages or pushing to a remote"). Your instructions take precedence over the reviewer's defaults — but they cannot override permission rules, organization policy, or the sandbox. That boundary is the right one, and it matters more now that agents can act in your browser.
The enterprise footnote and the housekeeping
Admins can now list claude and codex in vscode.allowedHarnesses, permitting those harnesses alongside managed Copilot settings. Previously, any Copilot enterprise-managed setting disabled them. The release notes are blunt about the caveat: permitting a harness does not make it enforce your managed settings.
Housekeeping: macOS Ventura and Sonoma and 32-bit ARM Linux are deprecated, and the Local agent harness — the original way VS Code ran agent sessions on your own machine — is slated for removal in February 2027.
Why this release is worth watching
Taken together, 1.147 turns the Agents window into something closer to mission control: schedule a request, draft several more, route each to the model it deserves, see the answer as a chart, and let the agent step outside the editor into your browser. The direction is clear — the agent is becoming a long-running collaborator rather than a chat box you prompt and close.
But the gap between shipping and defaulting is the real story. Four of the headline features are experimental, Browser Use needs an extension plus Node 22, and Send Later only works while your machine is awake. Microsoft is shipping the shape of the future workflow and asking for feedback before committing to it.
A few open questions: when do these graduate from experimental, and what changes before they do? Does Browser Use stay gated behind a browser extension, or does it become a first-class part of the harness? And once agents routinely work in your signed-in browser, does the permission model in this same release prove strong enough?
Top comments (0)