GitHub’s August 2026 Copilot changelog brings together changes shipped across VS Code versions 1.132, 1.133, 1.134, and 1.135. The updates focus less on a single model capability and more on the mechanics of working with AI coding agents: organizing multiple conversations, reconnecting to sessions from different clients, inspecting long transcripts, extending agents with plugins, and controlling terminal activity.
For developers using Copilot or other supported agent workflows, the practical effect is a more session-oriented experience. Several features are experimental or in preview, and availability can depend on the VS Code version, agent harness, platform, account configuration, or rollout status. VS Code 1.135 was released on August 26, 2026, with a gradual rollout.
What changed in the August releases
More control over long agent sessions
The Agents window now supports multiple chats arranged in horizontal or vertical groups. VS Code restores that layout when you return to the session, which makes it easier to keep separate tasks visible without repeatedly reopening conversations.
The experimental /btw command opens a side conversation while the primary chat continues running. The side conversation shares the primary chat’s context and prompt cache, making it suitable for asking a related question without interrupting an ongoing agent task.
Prompt-timeline navigation adds another way to review agent work. Developers can use the transcript gutter to move between prompts and related file changes. Chat also supports text search across the complete conversation transcript, with options for case-sensitive matching, whole words, and regular expressions.
For particularly long responses, experimental sticky scroll can keep the current prompt visible while you move through the output. The redesigned experience requires both chat.stickyScroll.enabled and chat.experimental.stickyScroll.enabled. VS Code 1.135 also uses a single-pane Agents window layout by default on desktop. Users who prefer the classic layout can disable sessions.layout.singlePaneDetailPanel and reload the window.
Sessions can move between clients
VS Code can show and continue recent Copilot or Claude agent sessions created in other applications. The chat.agentSessions.showExternal setting controls visibility, and the Sessions list includes an External filter.
The broader Agent Host architecture separates session ownership from the clients that display and control a session. A host can keep a session active when no client is connected, and the session can be reopened from another VS Code window or client. Local hosts communicate through message-port IPC, while remote connections use the Agent Host Protocol over WebSocket. VS Code documents code agent host as the command for running a standalone host.
This architecture is useful for developers who move between applications or VS Code windows during a longer task. However, the supplied documentation does not fully specify external-session eligibility, imported history, retention, concurrency limits, or every synchronization edge case. Those behaviors should not be assumed beyond the documented high-level support.
Portable Agent Plugins 1.0
Agent Plugins 1.0 defines an open packaging standard for portable skills and MCP servers across compatible AI-agent clients. The documented clients include GitHub Copilot in VS Code, GitHub Copilot CLI, and the GitHub Copilot app.
VS Code can install plugins from plugin marketplaces, and administrators or users can enable or disable plugin support with chat.plugins.enabled. Plugins may include hooks and MCP servers that execute code or expose tools on the developer’s machine. As a result, plugin installation should be treated as a code and permissions decision: review the manifest, hooks, MCP servers, contents, and publisher before enabling one.
The standard can make selected agent capabilities more portable, but it does not mean every client-specific behavior is identical. The available model, account, and organization controls for Copilot and Claude remain dependent on their respective configurations.
Voice, browser, and usage improvements
The integrated browser can now support selecting and annotating multiple HTML elements. It can also automatically reload changed local HTML files. Developers who frequently use browser-based agent workflows can configure the integrated browser as the default editor for HTML files through editor associations.
Built-in dictation is experimental and uses an on-device speech-recognition model by default. The first use downloads the default model unless a local package is installed. Dictation works in chat, the Agents window, editors, and terminals. Terminal dictation adapts speech for command-line input by removing ordinary punctuation, converting spoken symbol names, and adjusting capitalization.
The feature is available on Windows x64 and Arm64, macOS on Apple silicon, and Linux x64 and Arm64 with glibc 2.34 or later. It also works with remote workspaces because recognition runs on the local VS Code client. It is unavailable in VS Code for the Web, on Intel-based Macs, on 32-bit and Arm32 systems, and on musl-based Linux distributions such as Alpine Linux. The VS Code Speech extension is offered for those platforms.
Dictation requires AI features to be enabled and microphone access to be granted. User-level instructions can be stored in ~/.copilot/dictation.md, while trusted-workspace instructions use .github/dictation.md. Optional transcript cleanup sends transcript text, but not microphone audio, to a Copilot language model. Disabling cleanup keeps transcript processing local.
Chat response footers can show per-model input, cached-input, and output token usage when hovered. Session documentation also describes context usage and AI-credit consumption. These displays provide visibility into a chat turn, but they should not be treated as billing or quota meters without separate documentation.
The experimental /rubber-duck command is available in a Copilot Agent Host session. It asks a complementary model for a second opinion intended to surface missed details and edge cases. The exact model and model-selection behavior are not documented in the supplied sources, so its output should be treated as an experimental review aid rather than a substitute for testing or human review.
Sandboxing is available as an opt-in preview
Agent sandboxing uses OS-level isolation to restrict file-system and network access for terminal commands executed during agent sessions, including Copilot Agent Host sessions using the VS Code agent terminal integration. Sandboxed terminal commands can be automatically approved, while commands requiring additional permissions can prompt for execution outside the sandbox.
The scope is important. Sandboxing applies to shell subprocesses, not VS Code’s built-in file tools. Agent read, edit, and write operations continue to use VS Code’s permission system directly. The feature is therefore not a complete machine-isolation boundary. VS Code’s documentation recommends broader environment boundaries, such as development containers, when that level of isolation is needed.
On macOS, sandboxing uses Apple’s Seatbelt framework. Linux and WSL2 use bubblewrap and socat, and WSL1 is unsupported. Linux users must install the documented prerequisites. Network access can also be restricted with chat.agent.networkFilter, using allowed and denied domains. Allowing a domain can enable write actions against an external service, not merely read access.
In the VS Code 1.135 release notes, the default rollout for local-agent sandboxing was returned to 0% after a previous 50% rollout. The feature remains available as a UI opt-in and should be evaluated as preview software.
What developers and administrators should do
- Check the VS Code version. Review the relevant 1.132 through 1.135 release notes rather than assuming every August change belongs to 1.135. The 1.135 rollout was gradual, so earlier access may require Check for Updates or Insiders.
- Configure session visibility deliberately. Use
chat.agentSessions.showExternaland the External filter to control whether sessions from other applications appear in VS Code. - Review plugins before installation. Inspect plugin manifests, hooks, MCP servers, and publishers because these components can run code or provide tools locally.
- Test dictation against project data requirements. Confirm platform support, microphone permissions, model availability, and whether optional Copilot transcript cleanup is acceptable.
- Use sandboxing cautiously. Verify platform prerequisites and remember that built-in file tools are outside the sandbox. Do not rely on the preview as complete isolation.
- Keep normal engineering controls in place. Review generated changes, test agent-produced code, and apply appropriate approval and trust settings.
These releases make VS Code more capable as a control surface for persistent, multi-client agent work. The most significant operational change is the emphasis on sessions that can be organized, inspected, and potentially reopened across clients. Plugins, dictation, browser annotations, and sandboxing extend that workflow, but their preview status and configuration requirements make targeted testing important before wider deployment.
