VS Code 1.135 Makes AI Coding Sessions Easier to Continue

Visual Studio Code 1.135, released on August 26, changes how developers work with AI coding agents by making sessions less tied to a single editor window. You can now continue recent Copilot or Claude sessions started elsewhere, connect multiple VS Code windows to the same agent session, and ask a second model to review an agent's work. The update also exposes per-model token usage, giving developers a clearer view of what an AI-assisted coding task is actually consuming. :contentReference[oaicite:0]{index=0}

VS Code 1.135 treats an agent session more like a workspace

The biggest structural change is Agent Host, a dedicated process that runs agent harnesses separately from the individual VS Code windows connected to them. It is based on the Agent Host Protocol, an interface Microsoft is developing for synchronizing agent sessions across clients. Because the session is managed by the host rather than being tied to one editor window, multiple VS Code windows can connect to the same session. That matters when a developer wants one window for code, another for reviewing changes, or simply needs to reopen the work without rebuilding the conversation from scratch. :contentReference[oaicite:1]{index=1}

The distinction is subtle but useful: this is less about making the AI generate better code and more about making the work around an AI agent easier to manage. GitHub's August release summary describes the same direction across VS Code 1.132 through 1.135, including session organization, external-session continuation and Agent Host support. In other words, Microsoft is gradually turning the agent conversation from a temporary chat into something that can persist across parts of a development workflow. :contentReference[oaicite:2]{index=2}

You can continue a Copilot or Claude session from another app

VS Code 1.135 adds external agent sessions to the Sessions list. Recent Copilot or Claude sessions created in other applications can appear there, and selecting one lets you continue the conversation inside VS Code. The feature is controlled by the chat.agentSessions.showExternal setting and shows up to two recently updated external sessions by default. That means a developer can start an agent task in another client and move into the editor when the work reaches the point where direct code inspection becomes useful. :contentReference[oaicite:3]{index=3}

This also changes how developers should think about tool boundaries. Previously, moving from one AI coding application to another often meant losing context or manually explaining what had already happened. External session continuation reduces that friction, although the feature does not mean every AI tool or every session is automatically interchangeable. VS Code specifically documents support for recent Copilot and Claude sessions, so developers should not assume that an arbitrary agent conversation can be imported.

Rubber Duck gives an AI coding agent a second opinion

The most unusual addition is an experimental feature called Rubber Duck. In a Copilot Agent Host session, the /rubber-duck command asks a complementary model to examine the agent's work and surface missed details or edge cases. The idea is different from simply asking the same model to check itself: a second model can provide another perspective on the changes before the developer moves on. :contentReference[oaicite:4]{index=4}

There is an obvious limitation. Rubber Duck is an experimental review aid, not proof that code is correct. A second model can miss the same defect, misunderstand the requirements, or introduce its own incorrect criticism. Its useful role is closer to an additional review pass than an automated approval system, particularly for changes where a missed edge case would be expensive to discover later.

Token reporting makes AI coding costs easier to see

VS Code 1.135 also adds more detailed chat usage information. Hovering over a chat response footer reveals input tokens, cached input tokens and output tokens broken down by model for that turn. Tokens are the units used to process text and other model context, so this view gives developers a more direct picture of how much model processing a conversation is using. :contentReference[oaicite:5]{index=5}

That information becomes more useful as coding sessions grow longer. A short question may consume little context, while a long-running agent session can repeatedly carry project instructions, previous conversation and tool results into later turns. Seeing the breakdown does not automatically reduce the bill, but it gives developers evidence for deciding whether a particular model, workflow or review step is worth its cost.

The Agents window is now built around one side panel

The release also makes the single-pane side layout the default for the desktop Agents window. Session details and editors share one side pane, while the interface moves important information such as changes, pull requests, issues, browsers and artifacts into visible pills above the chat input. The layout can switch between side-by-side and inline diffs depending on available space, while developers who prefer the previous arrangement can disable the single-pane setting. :contentReference[oaicite:6]{index=6}

That redesign matters because an agent session is no longer just a conversation. The developer may need to inspect modified files, review a pull request, check browser output and follow the agent's progress without constantly changing views. Bringing those pieces closer to the conversation makes the editor better suited to supervising longer agent tasks rather than simply prompting an AI assistant.

What developers should actually take from the update

VS Code 1.135 is not primarily a new programming-language release or a performance upgrade. Its more significant change is organizational: AI coding work is becoming a persistent session that can move between applications and editor windows, while Agent Host provides the underlying process for keeping that work connected. The second-opinion feature adds another review layer, and token reporting exposes more of the resource cost behind the conversation. :contentReference[oaicite:7]{index=7}

For developers already using Copilot or Claude agents heavily, the practical starting point is simple: update VS Code, inspect the new Sessions workflow, and decide whether external-session continuation fits the way you work. If you experiment with Rubber Duck, treat its output as another reviewer's comments rather than a final verdict. And once agent sessions become long enough to consume significant context, the new token breakdown gives you something that was previously harder to see: where the model usage is actually going.