GitHub Copilot Dev Days

Key Concepts

• GitHub Copilot: An AI-assisted coding tool integrated across various platforms (CLI, VS Code, Visual Studio, JetBrains, Xcode, etc.).
• Agentic Development: The evolution of AI from simple autocomplete to autonomous agents capable of planning, executing tasks, and managing workflows.
• Copilot CLI: A terminal-based interface for AI-assisted development, supporting interactive, headless, and server modes.
• MCP (Model Context Protocol): A standard for connecting AI agents to external data sources, tools, and services.
• Skills: Specialized, community-contributed instructions and scripts that allow Copilot to perform specific tasks (e.g., code reviews, file manipulation).
• Autopilot & Bypass Modes: Advanced execution modes where agents perform tasks independently; "Bypass" ignores standard approval prompts, while "Autopilot" includes autonomous task completion and iteration.
• BYOK (Bring Your Own Key): A feature allowing developers to use their own API keys for various LLM providers (Anthropic, Mistral, etc.) within VS Code.
• Work IQ: An AI tool that integrates with meeting transcripts (e.g., from Microsoft Teams) to generate project plans and summaries.

---

1. Evolution of AI Development

The presentation outlines three "waves" of AI in software development:

• Wave 1 (2021): Smart autocomplete and "Ghost Text" acting as a passive pair programmer.
• Wave 2 (Current): Agentic development where humans provide specs, and agents execute workflows, with humans reviewing the output.
• Wave 3 (Future): Hybrid teams where humans and agents collaborate dynamically to build complex applications.

2. GitHub Copilot CLI: Features and Modes

The CLI is described as a "feature-rich agent" that treats the terminal as a first-class development environment.

• Headless Mode: Used for CI/CD automation. By using the -p (prompt) flag, developers can inject AI-generated content directly into scripts.
• Interactive Mode: Allows back-and-forth conversation. Users can use / commands to manage context, switch models, and view internal "thinking" steps.
• Server Mode: Exposes a JSON-RPC port, allowing external applications to programmatically leverage Copilot’s capabilities.
• Fleet Mode: Enables multiple agents to work on different parts of a task concurrently, with built-in orchestration to prevent race conditions.

3. VS Code and Visual Studio Integration

• Agent Window: A new top-level surface in VS Code designed to be "agent-first," allowing users to manage multiple sessions, workspaces, and remote connections (via SSH/tunnels) simultaneously.
• Remote Capabilities: Agents can run on remote machines, allowing developers to start a session on a desktop and continue it on a mobile device.
• Visual Studio 2026: Features include a new "Agent Preview" harness, test generation agents, and a profiler agent that simplifies complex performance analysis.

4. Methodologies and Frameworks

• Context Engineering: The practice of creating "grounding documents" (e.g., copilot-instructions.md) to provide agents with guardrails, design schemas, and project conventions.
• Human-in-the-Middle (HITM): A core philosophy where the human remains responsible for reviewing and validating agent output, especially given the non-deterministic nature of LLMs.
• Adversarial Validation (Rubber Duck Mode): Using two different model families (e.g., Sonnet and GPT) to critique each other’s work, which improves performance and reduces false positives.

5. Notable Quotes

• "We are right now in the middle stage of this wave two and moving on to wave three. We're going to have these hybrid teams of us and the agents working together." — Matt
• "The conversation now shifts from you building with AI to actually building AI using the CLI." — Julia
• "I'm a token-conscious kind of person and I like to watch what it's doing and stop it if it starts to go off the rails." — Wendy

6. Synthesis and Conclusion

The event emphasizes that GitHub Copilot is no longer just a coding assistant but a comprehensive ecosystem. By leveraging CLI tools, custom agents, and MCP servers, developers can automate repetitive tasks, perform complex code reviews, and even generate unit tests at scale. The key takeaway is the shift toward Agentic Development, where the developer acts as an orchestrator of multiple AI agents. Success in this new paradigm requires being explicit with instructions, utilizing planning modes to ensure predictable outcomes, and maintaining a "human-in-the-middle" approach to ensure code quality and security. Resources for further learning are available at gg.io/all-the-stuff and awesome-copilot.github.com.