Introduction

I have been building Agent Skills for this blog and other little projects for a month or so now, wiring up a draft-post-reviewer skill, has helped improve my posting and understanding, checking and validating my text and references. I run the VSCode insiders on my personal laptop but Agent Skills are now generally available, the new /init command can be used to scaffold workspace context from scratch.

Agent Skills Are Now Generally Available

Skills provide specialised capabilities, domain knowledge, and refined workflows that help the AI produce higher-quality output. Each skill folder contains instructions for specific domains (testing strategies, API design, performance optimisation, or in my case, blog post reviews ensuring content with the right front matter, tag taxonomy, and voice).

By default, VS Code looks for skills in .github/skills and .claude/skills in the workspace, or ~/.copilot/skills and ~/.claude/skills in the user home directory. Although I seen some posts about support for a more generalised .agents/skills coming soon.

Terminal Sandboxing

Terminal Sandboxing was a feature I was looking forward to trying, it is not something I have looked into at all yet due to it shipping with a macOS/Linux restriction, if this ends up being supported on Windows I look forward to trying it out more with my daily driver.

On supported platforms, terminal sandboxing restricts file system access to the workspace folder and lets you control network access to trusted domains only. When sandboxing is enabled, commands run without the standard confirmation dialog because they operate in a controlled environment.

MCP Apps

MCP Apps are something I have heard baout for a bit but not really go round to looking into fully. They allow MCP servers to render interactive UI elements in the client. The example in the VSCode docs is a Performance Profile graph.

I am watching this space because the potential new and better visualisations potentially for infrastructure status style dashboard elements inside chat, but I have not seen enough real-world examples yet. But hopefully that will come soon.

Integrated Browser (Preview)

The integrated browser in v1.109 replaces the old Simple Browser. The new browser is a proper embedded browser with persistent storage, full DevTools, keyboard shortcuts, and find-in-page. It supports configurable storage scope (global, workspace, or ephemeral).

There is also the “Add element to chat” feature which lets me select an element in the browser and send it directly to an agent for assistance.

Copilot Memory (Preview)

With the 109 release a preview feature was introduced, that provides a persistent memory store to save and recall important information between sessions. (enabled via github.copilot.chat.copilotMemory.enabled)

Combined with the existing instructions and skills, this creates a layered context system. Each layer serves a different purpose, and they complement each other rather than overlapping.

This is still a Preview feature, so I am not relying on it for anything critical yet, I look forward to playing with it further.

Agent Session Management

You can manage mulitple agents, all running concurrently either locally as an interactive chat, in the background or delegated to a cloud environment. Enabling you to switch between this agents is the session picker. This means you could use a model like Opus 4.5 to plan and analyse your requirements before offloading to Codex or Sonnet to do the implementation as background or cloud agents.

Claude Agent support is now in Preview. This integration uses Anthropic’s official Claude Agent SDK, sharing the same prompts, tools, and architecture as other Claude Agent implementations. I already use Claude Code to get a feel for the different tools.

Anthropic models also received improvements. The Messages API now supports interleaved thinking (reasoning between tool calls), and a tool search feature helps Claude discover the most relevant tools from a larger pool.

Key takeaways:

• Agent Skills are now officially GA. Custom skills and instructions files are now a supported, stable feature
• The /init command is the fastest way to bootstrap agent context for a new workspace, and it works alongside existing copilot-instructions.md files
• Terminal sandboxing is real but platform-limited. macOS and Linux get proper sandboxing today. Windows users (myself included) are still waiting
• MCP Apps open up interactive UI in chat responses, and custom registry base URLs enable enterprise MCP distribution through Azure DevOps feeds
• Copilot Memory with instructions together create a layered context system that scales from solo developers to enterprise teams

Further Resources

• VS Code January 2026 Release Notes (v1.109)
• MCP Apps SDK and Examples
• MCP Servers Documentation
• Agent Skills Documentation