What happened
GitHub says VS Code versions 1.132 through 1.135, shipped during August 2026, add a set of Copilot and agent-workflow features. The changes include side-by-side agent chats, side conversations, prompt navigation, portable agent plugins, model-provider switching, continuation of sessions created in other applications, and connections to the same session from multiple VS Code windows.
GitHub’s source is a release summary for VS Code versions 1.132, 1.133, 1.134, and 1.135, which it says shipped throughout August 2026. The announcement centers on Copilot-related agent sessions and workflows, with additional changes to chat, review, the integrated browser, and dictation. It is therefore a product update rather than a new model announcement or an independent evaluation of performance.
The largest group of changes concerns the Agents window. Users can arrange multiple chats side by side, open a side conversation with the /btw command while a primary chat continues, and use a prompt timeline to jump to earlier prompts and their related file changes. GitHub also says session details and editors can be viewed alongside chat in a single-pane layout, with diffs adapting to the available width.
The release adds several forms of session portability and configuration. GitHub says users can install agent customizations from plugins that follow the Agent Plugins 1.0 standard across VS Code and compatible agent clients. An experimental setting can open the Agents window without GitHub sign-in when Claude is configured with an API key. Claude sessions can switch between models from an Anthropic subscription and a Copilot subscription, while recent Copilot or Claude sessions created in other applications can be viewed and continued from VS Code. The Agent Host can also connect multiple VS Code windows to the same session.
Chat and review changes target long transcripts and generated edits. GitHub says users can search a complete chat transcript with case-sensitive, whole-word, and regular-expression options; keep the current prompt visible with sticky scroll; review rendered Markdown diffs through an experimental hybrid Markdown editor; switch between regular and diff editors from the breadcrumb bar; and view input, cached-input, and output token usage for each model by hovering over a response footer. The integrated browser can annotate multiple HTML elements for batch feedback, automatically reload local HTML files after changes, and become the default editor for HTML files. Built-in dictation adds multiple-language support through an on-device model, project-specific cleanup instructions, and shell-aware handling of terminal commands.
Why it matters
The update makes AI coding work more persistent, inspectable, and integrated into the editor. It also broadens the ways users can review agent-generated changes and give agents targeted feedback on local web pages. Those changes could reduce friction in longer coding tasks, while also making session boundaries, model choice, token usage, and permissions more important for developers and organizations to understand.
The update treats AI coding as an ongoing workflow rather than a sequence of isolated prompts. Persistent sessions, prompt navigation, visible file changes, and cross-window access can help developers reconstruct why an agent made a change and return to unfinished work. That is particularly relevant for projects where a coding task spans many files or requires several rounds of review.
The ability to continue Copilot or Claude sessions created elsewhere also points toward a more mixed tool environment. A developer may begin work in one application and continue it in VS Code, while model-provider switching can make subscription choice part of the workflow. GitHub’s announcement does not explain how context, permissions, billing, or data handling are managed when sessions cross applications or providers, so those operational details remain important unknowns.
The browser and review features could make agent-assisted web development more interactive. Selecting several HTML elements and sending targeted feedback in a batch may be more precise than describing page problems only in natural language. Automatic reloading of changed local HTML files could shorten the cycle between an agent edit and human inspection. These are workflow claims from GitHub; the source provides no independent measurements of time saved, defect reduction, or code quality.
The release also increases visibility into model usage by exposing input, cached-input, and output token counts for each model in a chat turn. That may help users understand consumption and compare how different models are used, but the announcement gives no pricing, quota, retention, or accounting details. On-device multilingual dictation may reduce the need to send audio elsewhere, yet GitHub does not specify supported languages, recognition accuracy, hardware requirements, or how project instructions are protected from unintended disclosure.
What to watch next
The key unknowns are how broadly the features are available, how experimental settings perform in practice, which agent plugins and model providers are supported, and how reliably sessions retain context across applications and windows. Developers should also assess the accuracy of agent edits, the privacy implications of shared session context, and whether on-device dictation works consistently across languages and shell commands.
Availability is the first practical question. Several capabilities are explicitly labeled experimental, including opening the Agents window without GitHub sign-in, the /rubber-duck command for a complementary model, and the hybrid Markdown editor. The source does not identify which features are enabled by default, which subscriptions are required, or whether all capabilities are available on every VS Code platform.
Developers should test the boundaries of shared context. Cross-application session continuation and connections from multiple VS Code windows can be useful, but they may also create confusion about which files, prompts, credentials, or instructions an agent can access. GitHub says the primary and side conversations share context and prompt cache, but the announcement does not describe isolation controls, conflict handling, or audit logs.
Agent plugins and model-provider switching will require attention to compatibility and governance. GitHub describes plugins following the Agent Plugins 1.0 standard and says Claude sessions can use models from Anthropic or Copilot subscriptions, but it does not list providers, plugin review requirements, or security safeguards. Organizations considering these features will need their own answers about extension trust, data movement, access control, and responsibility for agent-generated changes.
The quality of the review loop remains unverified by this source. Search, sticky scroll, adaptive diffs, browser annotations, and token counts improve visibility, but they do not establish that Copilot edits are correct or safe. The changelog contains no benchmark, error rate, user study, or independent assessment. Users should continue to inspect diffs, validate HTML behavior, review terminal commands, and treat agent output as proposed work rather than proof of correctness.
GitHub’s dictation claims also need practical testing. The source says the feature uses an on-device model, supports multiple languages, accepts user- or workspace-level cleanup instructions, and preserves shell syntax through shell-aware cleanup. It does not state how language detection behaves in mixed-language speech, how technical terms are handled, or what happens when cleanup instructions conflict with command syntax. Those details will determine whether the feature is dependable in real development environments.