Microsoft Ships Visual Studio Code 1.127 With GA Browser Tools and an Agent Organizing Overhaul
The latest update brings per-site permissions, subagent cost tracking, and a major push to tame AI session sprawl.
Microsoft has released Visual Studio Code 1.127, and the headline changes are all about taming the growing sprawl of AI agents while finally letting the integrated browser do what most developers already expected it could do. Browser tools for agents are now generally available, per-site permissions are here, and terminal commands on macOS and Linux now run in a sandbox by default.
VS Code has spent the last eighteen months pivoting hard toward AI-assisted workflows. 1.127 reads like a direct response to developer feedback on what happens when those workflows get messy. The Agents window now supports custom groups, drag-and-drop reordering, and quick actions to start or close out sessions. It's less about raw generation power and more about keeping your desk tidy when you've got six agents running at once.
Agents get groups, cost tracking, and CI banners
When you run several agent sessions simultaneously, the sessions list grows fast. You can now organize them into custom groups, collapse headers, and even drag a session onto a group to stash it. Select multiple sessions to move them as a block, or drop one onto the pinned section to keep it visible.
Chat input banners are probably the most time-saving addition. When a coding agent session has an open pull request, a banner sits right above the chat input. If CI checks fail, it shows how many broke and gives you a single click to launch an agent fix. Incoming PR comments trigger a similar banner with an "Address Comments" action that hands the feedback directly to the AI. No more alt-tabbing to a browser to check Jenkins.
Transparency on delegated work matters when you're burning through Copilot quota on parallel threads. Hover over any subagent section in a chat response to see exactly how many AI credits it consumed. The new /troubleshoot command also lets you diagnose agent behavior by analyzing session logs, though you'll need to pair it with #session and pick the active run in the Agents window.
Not exactly a clean room setup, but it works. The editor gutter now includes a feedback glyph you can hover over to drop comments on specific lines while reviewing agent changes. Pull request titles and descriptions also pull from the actual session context instead of defaulting to generic placeholders.
Browser tools go GA with per-site permission prompts
After months in preview, browser tools for agents ship as generally available. Agents can now open pages in VS Code's integrated browser, read console errors, take screenshots, and click through workflows to validate their own output. The feature required a per-site permission model to work inside the editor's security boundaries.
The camera, microphone, geolocation, Bluetooth, clipboard, and HID APIs are now supported. When a page requests a permission, VS Code fires a prompt exactly like a traditional browser would. You can manage active permissions through the Site Permissions menu item.
Microsoft says the per-session tab isolation and explicit page-sharing controls came straight from preview feedback. Enterprise admins can still disable the feature entirely or restrict which domains agents are allowed to reach via the ChatAgentNetworkFilter policy.
If you're wondering why this took so long, the answer is security. The integrated browser has historically been a locked-down space. Letting agents interact with full web APIs required a permission model that doesn't break existing trust boundaries. The long wait is over.
Terminal sandboxes and the Ollama deprecation
Approving every agent-invoked terminal command quickly becomes tedious. Starting with this release, commands run with network access blocked and filesystem access restricted on macOS and Linux. The agent only asks for approval when it actually needs to break out of the sandbox to run an elevated command.
The built-in Ollama provider is now deprecated. Microsoft says extension-based providers will ship new models and capabilities faster than a hardcoded provider ever could. If you were using the built-in option for local BYOK models, you'll need to grab the official Ollama extension and remove the deprecated provider. The migration is mostly frictionless, though self-hosted model purists might prefer a longer warning window.
Enterprise gets plain-file policy delivery
Administrators can now push managed Copilot settings via a plain JSON file at well-known OS paths: /Library/Application Support/GitHubCopilot/managed-settings.json on macOS, /etc/github-copilot/managed-settings.json on Linux, and %ProgramFiles%\GitHubCopilot\managed-settings.json on Windows. The file only takes effect when native MDM or account-based enterprise settings aren't present, so it slots cleanly into existing imaging pipelines.
It's a solid polish release rather than a paradigm shift. The terminal sandboxing and cost transparency are the kind of improvements you don't realize you needed until they're there. Multi-chat sessions finally get tab-level management, and the responsive sidebar behavior cuts down on accidental clicks when working on a narrow window. Though, forcing Ollama users to switch extensions might ruffle a few feathers among teams that haven't fully audited their local model supply chain yet.
VS Code 1.127 is rolling out gradually to the stable channel. Hit Check for Updates inside the editor to grab it now. Insiders can download the nightly build for immediate access to the latest patches. Head here for the full release notes and the step-by-step guide to building web apps with the new browser agent tools.
