Program Your Agents

GitHub Copilot's February release introduces powerful agent automation and control mechanisms:

• Lifecycle hooks: Run custom code at key agent events to enforce policies, auto-lint code, or block commands before execution
• Conversation forking: Branch from any checkpoint to explore alternative paths without losing your starting point
• Auto-approval controls: Toggle global auto approval with /autoApprove or /yolo commands in chat, with optional terminal sandboxing for bounded execution
• Real-time steering: Send follow-up messages while agents work to redirect their approach without waiting for completion

Extend Agent Capabilities

New extensibility features allow developers to customize and expand agent functionality:

• Agent plugins: Install prepackaged skill bundles, tools, hooks, and MCP servers from the Extensions view (experimental)
• Skills as slash commands: Trigger agent skills directly from chat, including those from custom extensions
• Agentic browser tools: Agents can now navigate, click, screenshot, and verify changes in the integrated browser (experimental)
• Dynamic customization: Generate reusable prompts, skills, agents, and hooks from chat using /create-* commands

Additionally, Copilot CLI is now natively integrated into VS Code with diff tabs, trusted folder sync, and right-click code snippet sharing.

Intelligent Context Management

Agents now maintain better awareness across sessions with shared memory and automatic context optimization:

• Cross-tool memory: Agent knowledge persists across the Copilot coding agent, CLI, and code review for cumulative understanding
• Plan persistence: Plans survive across conversation turns and automatic compaction, building on previous work
• Explore subagent: Lightweight models handle fast, parallelized codebase research, providing specific file and code path references
• Manual context control: Trigger compaction with /compact and guide retention with natural language (e.g., /compact forget about all variants, except the rust version)
• Large output handling: Tool outputs are written to disk rather than stuffed into context, preserving important details during compaction

Long-distance next edit suggestions now predict edits anywhere in your file, improving code navigation efficiency.