Rubber Duck Thursdays!

Key Concepts

• Agentic Coding: A workflow where AI agents (like GitHub Copilot) autonomously gather context, execute tasks using specific tools, and verify results.
• Agentic Loop: The iterative process of prompting, context gathering, tool execution, and result verification.
• GitHub Copilot Chat: An AI-powered interface within VS Code that supports context-aware coding and agentic interactions.
• Reasoning Effort: A configuration setting in VS Code that allows users to adjust the model's "thinking" depth (Low, Medium, High) based on task complexity.
• Skills: Modular, task-specific instructions (defined in skill.md files with YAML headers) that guide Copilot to perform predictable workflows, such as email management or PowerPoint creation.
• Custom Agents: Specialized AI personas (defined in agent.md files) that allow for role-specific behavior (e.g., a "Reviewer Agent") distinct from global repository instructions.
• Copilot CLI: A command-line interface for Copilot that supports "YOLO mode" and remote, read-only sessions for monitoring long-running tasks.

---

1. The Agentic Loop and Workflow

Marlene explains that effective agentic coding relies on a structured loop. Instead of treating AI as a simple chatbot, developers should:

• Gather Context: Use features like referencing specific GitHub issues or pull requests directly within the chat to ground the AI's responses.
• Utilize Tools: Leverage specialized skills and agents to perform actions rather than just generating text.
• Verify Results: Use the agent to perform the task and then review the output, iterating as necessary.

2. Managing Reasoning Effort

A significant update in VS Code is the ability to control Reasoning Effort.

• Low Reasoning: Best for quick tasks, presentations, or simple code snippets where speed is prioritized over deep analysis.
• Medium Reasoning: The recommended default for day-to-day development tasks.
• High Reasoning: Reserved for complex research, architectural planning, or debugging intricate logic where the model needs to "overthink" to ensure accuracy.

3. Customizing Behavior: Skills vs. Agents

Marlene distinguishes between two ways to extend Copilot’s capabilities:

• Skills (skill.md): Focused on tasks. These are used for specific, repeatable actions (e.g., using the Graph API to draft emails or using an Anthropic-style PowerPoint skill to generate slide decks). They are highly predictable and follow a defined workflow.
• Agents (agent.md): Focused on behavior. These define how the AI acts (e.g., a "Reviewer Agent" that adopts a specific persona). Unlike global copilot-instructions.md files, which affect the entire repository, custom agents allow for context-specific behavior that can be toggled on or off.

4. Real-World Applications

• Presentation Design: Marlene uses Copilot CLI and Claude Opus 4.6 to generate HTML/CSS for presentation slides. By providing examples of preferred styles, she allows the agent to iterate on designs (e.g., changing button colors or adding animations) rapidly.
• Automated Testing: The Playwright Agent (installed via npx) is highlighted as a powerful toolset consisting of three distinct agents: one for planning, one for generating tests, and a "healer" agent that automatically fixes broken tests.
• Email Triage: Using an MCP (Model Context Protocol) server, Marlene demonstrates how a custom skill can interface with the Graph API to automate email drafting and management.

5. Notable Quotes

• "Vigilance alone doesn't fix problems; community, however, helps us see them in a larger focus." (Attributed to Alistister in the chat).
• "I like to think about the skills as more of a task thing... If I want that model to behave in a certain way, then I'm going to create an agent for that." (Marlene).

6. Synthesis and Conclusion

The session emphasizes that the future of AI-assisted development lies in moving beyond simple code completion toward agentic workflows. By utilizing custom skills for specific tasks and agents for specialized roles, developers can create a more predictable and efficient coding environment. The key takeaway is to treat the AI as a tool-using collaborator that requires proper context (issues, PRs, and clear instructions) to function at its highest potential. Marlene encourages developers to experiment with the skills and agents folders in their repositories to tailor Copilot to their specific professional needs.