Program Your Agents with Lifecycle Hooks

GitHub Copilot agents now support programmatic hooks that run at key lifecycle events, giving developers explicit control over agent behavior and policy enforcement. Key capabilities include:

• Automate with hooks: Enforce policies, auto-lint, or block commands before execution at designated agent lifecycle points
• Fork conversations: Branch from any checkpoint to explore alternative paths without losing your starting point
• Auto-approve from chat: Toggle global auto-approval with /autoApprove or /yolo commands, paired with terminal sandboxing to safely contain agent execution
• Queue and steer: Send follow-up messages while agents are working to redirect their approach without waiting for completion

Extend Agents with Plugins and Skills

The release introduces new extensibility mechanisms for customizing agent capabilities:

• Agent plugins: Install prepackaged bundles of skills, tools, hooks, and MCP servers from the Extensions view (experimental feature)
• Skills as slash commands: Trigger agent skills directly from chat, including those from third-party extensions
• Agentic browser tools: Agents can now drive the integrated browser to navigate, click, screenshot, and verify their own changes (experimental)
• Generate customizations from chat: Use /create-* commands to generate reusable prompts, skills, agents, and hooks directly from conversations
• Copilot CLI integration: Copilot CLI now ships natively in VS Code with diff tabs, trusted folder sync, and right-click code snippet sending

Smarter Context Management and Memory

The update significantly improves how agents manage context and knowledge across sessions:

• Shared agent memory: Knowledge persists and builds across the Copilot coding agent, Copilot CLI, and code review
• Plan memory: Plans persist across conversation turns and compaction, enabling agents to build on prior work rather than restarting
• Explore subagent: Lightweight models handle fast, parallelized codebase research, allowing the main agent to reference specific files and code paths
• Context compaction: Automatic compaction with manual override via /compact command (e.g., /compact forget about all variants, except the rust version)
• Large output handling: Tool outputs exceeding context limits are written to disk to preserve important details during compaction
• Long-distance edit suggestions: Next edit predictions now work anywhere in a file, not just local regions

Additional Improvements

The release also includes general VS Code productivity improvements: Kitty graphics protocol support for terminal images, a redesigned model picker with search and rich details, improved chat accessibility for screen readers and keyboard navigation, and AI coauthor attribution for commits containing AI-generated code.