Workshop Summary: GitHub Copilot Agents, Customization, and MCP

Facilitators: Julia Muiruri, Bryan, Kelvin Mbuco, Agustin, Eldrick, Lilian Njuguna
Date/Time: Session covering agent workflows, UI updates, customization, and external tooling integration

Copilot Agent Workflows

• Specialized Agents:
  • The workflow uses multiple specialized agents rather than a single agent handling all tasks.
  • Example workflow:
    1. Incident/OCE Agent: Handles incident analysis, error summaries, network issues, logs, telemetry. Also coordinates workflow.
    2. Coding Agent: Takes over for code changes, PR creation, testing, deployment preparation.
    3. CICD: Automated deployment is separate; not an agent per se.
• Delegation Philosophy:
  • Assign specific tasks to specialized agents to simplify workflow and ensure clarity.
  • Humans still handle managerial/strategic tasks while AI handles repeatable, well-defined actions.
• Practical Demonstration:
  • Adding a delete button to a recipe app using Copilot Agent Mode.
  • Agents can analyze code, create actionable to-dos, implement changes, and run terminal commands (requires user approval).
  • Features like hidden terminal and checkpoint restore provide safety and visibility into agent actions.

Interaction with UI Elements

• Element Selection Feature:
  • Users can select UI components directly in the app.
  • Copilot Agent receives context about selected elements for implementing functionality (e.g., adding functionality to cards or updating recipe elements).
  • This avoids long or ambiguous prompts.
• Best Practices in Prompting:
  • Follow the 3S Rule: Short, Simple, Specific prompts.
  • Step-by-step instructions preferred over large, single-step prompts for better control.

Customization and Context

• Custom Instructions:
  • Generic Instructions: Applied to the entire project; defines overall coding style, company conventions, libraries to use.
  • File-Specific Instructions: Applied to specific file types or folders (e.g., .CSS files in a public directory).
  • Generated Instructions: Copilot can analyze an existing codebase and create instructions based on existing standards.
• Best Practices:
  • Clean up your codebase before onboarding Copilot to ensure best practices are propagated.
  • Leverage community-curated instruction sets from repositories such as Awesome Copilot Instructions for various languages and frameworks.

MCP (Model Context Protocol)

• Purpose:
  • Extends Copilot’s capabilities by giving agents access to external tooling, data, and services.
  • Example: Analyzing logs in Kusto or performing GitHub actions.
• How It Works:
  • MCP servers define available tools and actions.
  • Copilot can invoke these tools for tasks (e.g., GitHub fork creation, fetching documentation).
  • Provides more accurate, contextual, and real-time information compared to relying solely on training data.
• Practical Example:
  • Using GitHub MCP server to create a fork.
  • Using Microsoft Docs MCP server to generate a 4-week learning plan based on official modules.
• Security Considerations:
  • Tool poisoning and hidden malicious instructions are active concerns.
  • Use vetted MCP servers from the official MCP registry as a first layer of protection.
  • Ongoing discussions in VS Code community and Discord office hours for advanced threat mitigation.

Key Features Demonstrated

 Participant Assignments / Challenges

1. Complete Section 3: Add functionality to app cards (bonus: change UI creatively).
2. Test Copilot Agent Mode by adding delete functionality for recipes.
3. Explore Section 4: Custom Instructions and MCP setup.

Takeaways

• Specialized agents + clear delegation improve efficiency.
• Step-by-step, explicit prompts help Copilot stay on track.
• Custom instructions ensure Copilot aligns with project standards.
• MCP enables safe integration with external tools for more advanced automation.
• Security awareness is critical when using MCP servers; stick to vetted sources.