"When you input a chat prompt using agent mode in VS Code, the model can invoke various tools to perform tasks such as file operations, accessing databases, or retrieving web data," the dev team said. "This integration enables more dynamic and context-aware coding assistance."

The team also introduced several new built-in tools for Agent Mode, providing a variety of AI-driven features to enhance productivity and task execution.:

• Thinking Tool (Experimental): Allows the model to "think" between tool calls to improve performance on complex tasks, inspired by Anthropic's research.
• Fetch Tool: Lets you include content from publicly accessible webpages in prompts, caching the page data locally for repeated use (without JavaScript support and for non-authenticated pages only).
• Usages Tool: Combines "Find All References," "Find Implementation," and "Go to Definition" to help explore and analyze code more effectively, especially useful for refactoring tasks.

There were also many more minor tweaks, fixes, observations and improvements to the Copilot Chat experience, including:

• Create a New Workspace with Agent Mode (Experimental): Scaffold a new VS Code workspace in agent mode to initialize, configure, and launch development environments like extensions or MCP servers with necessary dependencies and settings. Setting: github.copilot.chat.newWorkspaceCreation.enabled
• VS Code Extension Tools in Agent Mode: Extension tools contributed to VS Code's language model tools API are now available in agent mode. These tools can be easily installed and managed through the Extension Marketplace. See the language model tools extension guide for building your own.
• Agent Mode Tool Approvals: You can now approve the use of tools and terminal commands in agent mode, with options to remember approval on a session, workspace, or application level. A new experimental setting, chat.tools.autoApprove, allows for auto-approving all tools without confirmation.
• Agent Evaluation on SWE-bench: VS Code's agent achieves a 56.0% pass rate on the SWE-bench evaluation using Claude 3.7 Sonnet, reflecting improved agent capabilities with new tools and optimized prompts.
• Unified Chat View: The separate "Chat" and "Copilot Edits" views are now merged into a single "Chat" view, allowing you to switch between "Ask," "Agent," and "Edit" modes for more flexible AI-assisted coding and task management. New features include switching modes mid-conversation, restoring edit sessions, and running multiple agent sessions.
• Bring Your Own Key (BYOK) (Preview): Copilot users can now bring their own API keys for external providers like Azure, Anthropic, Gemini, OpenAI, Ollama, and Open Router, allowing access to new models as soon as they are released.
• Reusable Prompt Files: The chat.promptFilesLocations setting now supports glob patterns, allowing flexible inclusion of files (e.g., all .prompt.md files in a workspace). Improved prompt file editing includes autocompletion and error squiggles for invalid references.
• Alignment with Custom Instructions: The .github/copilot-instructions.md file now supports enhanced language features and can be treated like other reusable .prompt.md files, with better link resolution and handling.
• User Prompts: The new "Create User Prompt" command lets you create custom prompts stored in the user data folder, with synchronization across machines like code snippets or settings. These prompts can be synchronized via the Sync Settings menu.
• Improved Vision Support (Preview): Copilot Vision now supports image attachments in chat, including the ability to drag and drop images from browsers with proper file extensions (.jpg, .png, .gif, .webp, .bmp).

While GitHub Copilot AI, particularly the Chat functionality, received the most attention in the release, work also continued on other areas like code editing, accessibility, source control and more.