Live
Self‑Hosted Void Platform Lets Teams Govern AI‑Built Apps on CloudflareRethinking Agent Evaluation: Practical Lessons from the Google Cloud Developer PluginTaming Semantic and Embedding Drift in Production Agentic SystemsLLM autonomous hacking exposes gaps in model sandboxing and credential handlingGitHub Actions GA Xcode 27 Runner on Apple SiliconSeparating Orchestration and Compute for GPU‑Intensive Biotech AgentsAgent Quality Evaluation at Scale with Amazon Bedrock AgentCoreAI agents on Kubernetes: handling 20‑fold request spikes and new runtime optionsSelf‑Hosted Void Platform Lets Teams Govern AI‑Built Apps on CloudflareRethinking Agent Evaluation: Practical Lessons from the Google Cloud Developer PluginTaming Semantic and Embedding Drift in Production Agentic SystemsLLM autonomous hacking exposes gaps in model sandboxing and credential handlingGitHub Actions GA Xcode 27 Runner on Apple SiliconSeparating Orchestration and Compute for GPU‑Intensive Biotech AgentsAgent Quality Evaluation at Scale with Amazon Bedrock AgentCoreAI agents on Kubernetes: handling 20‑fold request spikes and new runtime options
GitHub

August 2026 VS Code updates tighten Copilot agent workflows and add multilingual dictation

AI SummaryPowered by AI

VS Code 1.132‑1.135 introduce side‑by‑side chat panes, side‑conversation commands, portable agent plugins, and multilingual dictation. These changes give AI‑engineers and platform teams finer control over agent sessions, improve traceability of generated code, and add operational considerations for plugin management and language handling.

VS Code versions 1.132 through 1.135, released in August 2026, add a set of UI and configuration tweaks that reshape how Copilot and Claude agents are managed inside the editor. The updates focus on arranging multiple chat windows, launching side‑conversations that share context, installing portable agent plugins, and extending dictation to multiple languages—all of which affect the way AI‑engineers, cloud/platform engineers, and SREs interact with generated code and the surrounding workflow.

Agent session organization

The Agents window now supports side‑by‑side chat groups, allowing users to keep several conversations visible in either horizontal or vertical splits. The layout persists when returning to a session, which reduces the friction of comparing alternative suggestions. A new /btw command opens a side‑conversation that inherits the primary chat’s prompt cache while the original continues running, enabling parallel exploration without losing context.

Experimental support for opening the Agents window without a GitHub sign‑in is available when Claude is configured with an API key, which can simplify onboarding in isolated environments. Users can also switch model providers on the fly between Anthropic and Copilot subscriptions, and they can resume external Copilot or Claude sessions from other applications via the Sessions list. The Agent Host can now connect multiple VS Code windows to the same session, facilitating collaborative debugging or pair‑programming.

Chat navigation and review

Long conversations become searchable with a full‑transcript text search that respects case, whole‑word, and regex options. A “sticky scroll” feature pins the current prompt while scrolling, preventing loss of context in extensive threads. The transcript gutter now includes a prompt timeline control for jumping directly to a specific prompt and its associated file changes.

Diffs are now rendered alongside a hybrid Markdown editor, allowing inline editing of generated documentation. Users can toggle between regular and diff editors from the breadcrumb bar, and terminal output within chat reflows automatically when the pane is resized. Hovering over a response footer reveals token usage broken down into input, cached input, and output for each model, giving visibility into cost and performance.

Integrated browser enhancements

The built‑in browser can annotate multiple HTML elements on a page, letting agents apply batch UI feedback. Local HTML files automatically reload in the browser when they change on disk, providing immediate visual confirmation of agent edits. An editor association setting lets the integrated browser become the default editor for HTML files, streamlining a preview‑first workflow.

Dictation improvements

Dictation now runs on an on‑device model that can detect or be set to a specific language, expanding voice‑driven coding beyond English. Users can add project‑level or workspace‑level cleanup instructions so that spoken input respects project terminology and formatting conventions. For terminal commands, a shell‑aware cleanup step preserves command syntax, avoiding the insertion of literal punctuation that previously required manual correction.

Related CloudNinjas coverage: AI engineering.

What This Means For Practitioners

These UI‑level changes do not alter the underlying Copilot or Claude APIs, but they shift operational responsibilities. Teams should evaluate the experimental “no‑sign‑in” mode for isolated CI environments and assess the security posture of portable agent plugins before adoption. The ability to switch model providers mid‑session may affect budgeting and token‑usage monitoring, so integrating the token‑usage hover data into existing cost‑tracking pipelines is advisable. Multilingual dictation introduces new language‑specific tokenization considerations; validation of generated code remains essential, especially when voice input is used in production scripts. Finally, the side‑by‑side chat layout and shared prompt cache encourage collaborative debugging, but they also require clear ownership of session state to avoid unintended overwrites in shared environments.

Originally published atGitHub Changelog