Rubber Duck Thursdays - Let's build with GitHub Copilot SDK

Key Concepts

• GitHub Copilot SDK & CLI: A new toolkit enabling developers to integrate Copilot’s agentic capabilities directly into applications, moving beyond simple code completion.
• Context Engineering: The focus is shifting from crafting perfect prompts to effectively managing the context provided to AI models, including tool access and conversation history.
• AI Augmentation: Copilot is positioned as a collaborative tool to augment engineering expertise, not replace it, accelerating development and freeing engineers for higher-level tasks.
• Model Context Protocol (MCP): A standardized protocol for connecting AI agents to tools and services, exemplified by integration with Work IQ and GitHub.
• Agentic Workflows: Utilizing custom agents within the SDK to orchestrate complex workflows from specification to code.

Introduction to the GitHub Copilot SDK & Flight School (Part 1)

The stream began with an introduction to the newly released GitHub Copilot SDK, described as a wrapper around the Copilot CLI. This allows developers to embed Copilot’s agentic capabilities directly into their applications. A key demonstration project, “Flight School,” was presented as a practical example. Flight School is a learning application that personalizes recommendations based on a user’s GitHub profile and skills. The hosts emphasized a shift in focus from “prompt engineering” to “context engineering” – optimizing the information provided to the AI, including tool access and conversation history. The Model Context Protocol (MCP) was introduced as a standardized way to assemble AI agents and connect them to tools and services, with Work IQ (accessing Teams and Outlook data) as an example. The concept of “progressive disclosure” was also highlighted, suggesting gradually revealing information to the AI to optimize the context window. The hosts noted the rapid iteration of the Copilot CLI and SDK, with frequent releases.

Demonstrating Copilot’s Capabilities & Integration (Part 2)

The second segment focused on a live demonstration of the Copilot CLI and SDK, showcasing its capabilities beyond basic code completion. The demonstration utilized the Monaco editor, the same editor powering VS Code, and centered around a challenge-solution-evaluation loop. Copilot was presented with a coding challenge, generated a solution, and then was used to evaluate the solution’s completeness. This process highlighted that Copilot doesn’t always achieve 100% accuracy, necessitating human review. The hosts emphasized viewing Copilot as a collaborator, assisting with tedious tasks like scaffolding and boilerplate code. The distinction between the CLI (for direct interaction) and the SDK (for embedding functionality) was clarified. Copilot’s ability to leverage external tools via “tool calls” was demonstrated, using both the octokit package (for GitHub interaction) and the GitHub MCP server. A practical example showed Copilot automatically generating and submitting bug reports to GitHub, including steps to reproduce. The hosts also demonstrated prompting Copilot to refactor code for maintainability, creating folders and tests.

Key Perspectives & Analogies

Throughout both segments, the hosts consistently argued that AI, specifically Copilot, should augment human engineers, not replace them. Engineering principles like testing and review remain crucial. They described the evolving role of the engineer as becoming more of a “director” or “orchestrator,” guiding AI assistants to realize their vision. Humorous analogies were used to illustrate Copilot’s capabilities: comparing it to a “toddler” – powerful but requiring careful guidance and supervision – and referencing the nostalgic experience of interacting with a text-based “Mud” (Multi-User Dungeon). Chris Noring stated, “GitHub Copilot completely has changed the game for us all… We’re able to build things in minutes and learn things.”

Technical Details & Terminology

Key technical terms and concepts discussed included: the GitHub Copilot SDK and CLI, the MCP (Microsoft Copilot Platform), Work IQ, the Context Window, Progressive Disclosure, YOLO Mode, the Monaco Editor, Agentic Experience, Tool Calls, Octokit, and the concept of an Agent within the SDK. The importance of providing Reproducible Steps (Repro) for bug reporting was also highlighted.

Conclusion

The Rubber Duck Thursdays stream provided a comprehensive overview of the GitHub Copilot SDK and CLI, demonstrating its potential to significantly accelerate developer workflows. The emphasis on context engineering, AI augmentation, and the evolving role of the engineer suggests a future where developers focus on orchestration and high-level problem-solving, leveraging AI as a powerful collaborative tool. While acknowledging the need for human oversight and the limitations of current AI models, the stream presented a compelling vision for the future of software development.