VS Code 1.128: Parallel Agent Sessions Are Here Now

Microsoft released VS Code 1.128 on July 8, introducing parallel agent sessions that allow multiple Claude conversations to run simultaneously within a single window, along with subagent monitoring and generally available browser tools for GitHub Copilot. The update also graduates Copilot Vision to general availability and adds OS-level keyboard shortcuts for the Agents window, marking a shift toward AI-integrated workflow rather than bolted-on features.

VS Code 1.128 dropped on July 8 and it draws a line. Microsoft is no longer treating VS Code as an editor with AI bolted on the side. The headline feature — multi-chat agent sessions — lets you run parallel Claude conversations inside a single window, fork threads at any turn, and monitor subagents in read-only view. Browser tools are now generally available https://github.blog/changelog/2026-07-01-browser-tools-for-github-copilot-in-vs-code-are-generally-available/ and on by default. Copilot Vision images and PDFs in chat graduates from preview to GA. And OS-level keyboard shortcuts let you summon the Agents window without switching apps. This release is about workflow, not features. Parallel Agent Sessions: The Shift That Changes How You Work The core change in 1.128 is straightforward to describe and significant to absorb. Claude agent-host sessions in the Agents window now support multiple chats per session. Each chat keeps its own history, title, and model selection. Peer chats stay grouped under the parent session rather than spawning independent top-level sessions, and they restore together after restart. In practice, this means you can run your implementation work in one chat, your test generation in another, and your documentation in a third — all under the same session umbrella, simultaneously. If you want to explore a different approach midway through, you fork a chat from an earlier turn. Neither thread loses its context. You switch between them with ⇧⌘ and ⇧⌘ . Before 1.128, AI coding was fundamentally serial. You waited for the agent to finish one task before starting the next, or you juggled separate sessions with no shared context. The model now more closely mirrors how engineering teams actually work: multiple people on different parts of the problem at once. The difference is that your “team” runs inside a single window. Quick chats ⌘K ⌘N let you ask one-off questions without opening a workspace at all — useful when you need a fast answer without spinning up a full session. Subagent Monitoring: Watch Without Interfering When an agent delegates work to subagents — say, you have asked it to refactor an entire module and it spawns workers for individual files — those worker transcripts now appear as read-only peer chats. They are hidden from the tab strip by default; you open them from the Conversations menu, the running-subagents chip, or an inline subagent pill in the parent transcript. The view-only nature is intentional. You can see exactly what each worker is doing without accidentally steering or interrupting it. Complex orchestration patterns, which previously happened invisibly, are now observable. This matters less for small tasks and a great deal for the kind of large-scope refactors where you genuinely need to know which file the agent is touching right now. Browser Tools: The Agentic Loop Closes Browser tools for GitHub Copilot went GA on July 1 and 1.128 enables them by default for all paid Copilot subscribers. Agents can now navigate to a URL, click through a live application, read page content and console errors, take screenshots, and verify that code actually works — without you leaving the editor. This is what “closing the agentic loop” means. Previously, an agent could write code but could not confirm the result. Now the workflow is complete: write code, run it, have the agent open the browser and verify the feature works, then report back. The agent checks its own work. Privacy controls are solid. Your existing tabs are private by default; agents only access pages you explicitly share. Agent-opened pages run in isolated sessions with no access to your cookies or stored data. Parallel agents maintain separate browser contexts. High-risk browser capabilities camera, mic, location require explicit per-site approval. Enterprise admins get domain filtering and workspace trust controls on top of that. Copilot Vision and System-Wide Shortcuts Image and PDF support in Chat is now generally available. Attach files by pasting, dragging, or dropping — or let agents access images directly through tool calls. The use cases are well-established: paste a UI screenshot of a bug and let the agent locate the relevant component; drag in a design mockup and get matching implementation code. The full June 2026 Copilot changelog https://github.blog/changelog/2026-07-08-github-copilot-in-visual-studio-code-june-2026-releases/ covers the complete feature list. OS-level keyboard shortcuts are the smaller-but-useful addition. Add "systemWide": true to any keybinding in keybindings.json and that shortcut fires even when VS Code is not in focus. The main use case: summon the Agents window from wherever you happen to be — browser, Slack, terminal — without alt-tabbing first. // keybindings.json — system-wide shortcut to open Agents window { "key": "ctrl+shift+a", "command": "workbench.view.extension.claude-agent", "systemWide": true } The Verdict VS Code 1.128 is not a features release — it is a workflow release. The full release notes https://code.visualstudio.com/updates/v1 128 list more additions, but the parallel sessions model is the one that changes the mental model for AI-assisted development. You are no longer waiting on one agent at a time; you are orchestrating a small team. Combined with browser-verified output and visible subagent work, the IDE is doing something qualitatively different from what it did six months ago. Update, open the Agents window, and start thinking in parallel.