Rubber Duck Thursdays

Key Concepts

• GitHub Copilot CLI: A command-line interface tool for interacting with AI agents to perform development tasks.
• MCP (Model Context Protocol): A standard for connecting AI agents to external tools, data sources, and services.
• Fleet Orchestration: A feature allowing multiple AI agents to run in parallel to handle different parts of a project simultaneously.
• Auto Mode: A CLI setting where Copilot intelligently selects the best model based on service health and availability, often providing a cost-effective usage rate.
• Telemetry: Usage data collected by GitHub to improve product features, with options for users to inspect or opt out.
• Self-Healing/Agent Loop: The ability of an AI agent to diagnose errors, iterate on its own logic, and apply fixes during task execution.

---

1. Main Topics and Announcements

The session focused on the latest updates from the GitHub Changelog (as of April 23rd) and a live demonstration of the Copilot CLI.

• Jira Integration Enhancements:
  • Custom Agents: Users can now assign specific Copilot agents to Jira tickets.
  • Custom Fields: Copilot can read Atlassian custom fields (e.g., acceptance criteria) to provide better context.
  • Branching Rules: The agent now respects existing branch naming conventions defined in Jira.
  • Custom Instructions: Users can define repetitive configurations in an instructions file to be applied automatically.
• Copilot Code Review Metrics: New parameters in the Usage Metrics API allow enterprise admins to track "active" vs. "passive" users. An active user is defined as someone who manually requests a review or applies a suggestion.
• Bring Your Own Model (BYOM): Business and Enterprise users can now connect Copilot to their own API keys from providers like Anthropic, Gemini, OpenAI, or local models via Ollama/Foundry.
• Copilot Plan Changes: GitHub has paused new sign-ups for Pro, Pro Plus, and Student plans due to high abuse of free trials. Additionally, tighter session and weekly token limits have been introduced, and "Opus" models have been removed from the standard Pro subscription.

---

2. Step-by-Step Workflow: AI-Driven Development

The host demonstrated a professional workflow using the Copilot CLI:

1. Data Transformation: Used the CLI to convert a raw product_data.json file into an organized Excel spreadsheet with visualizations.
2. Research Phase: Used the /research command to have the agent study the documentation of a design tool (Pencil) to understand its capabilities before implementation.
3. Planning Phase: Used /plan mode to generate a structured development roadmap, which was saved as plan.md for future context.
4. Design Generation: Connected the agent to the Pencil MCP server to generate UI designs based on the research and the product data.
5. Fleet Orchestration: Used the /fleet command to run two agents in parallel: one to review design readiness and another to build the application (Electron app).

---

3. Key Arguments and Perspectives

• Context is King: The host argued that AI models are only as powerful as the context they can access. By bringing models close to where data resides (e.g., Jira, local files), developers achieve higher efficiency.
• Transparency: Regarding telemetry, the host emphasized that GitHub provides full visibility into what data is collected. Users can use environment variables to preview the JSON payload sent to GitHub before deciding to opt out.
• Tool Selection: When asked about the "best" IDE or CLI, the host suggested that there is no universal answer. The best tool is the one that fits the developer's existing rhythm and integrates seamlessly with their specific stack.

---

4. Notable Quotes

• "Models are only powerful given the context that they have access to. If you can bring these models close to where your data sits... these models will easily just pick up on how you're already working."
• "The biggest favor you can do for yourself right now is just experiment. Just try everything out. Pick out what doesn't work... and whatever works, just work on improving it."

---

5. Synthesis and Takeaways

The session highlighted a shift toward agentic workflows where the CLI acts as an orchestrator rather than just a code-completion tool. By utilizing MCP servers and Fleet orchestration, developers can automate complex, multi-step tasks—from data analysis to UI design and application building—within a single terminal environment. The key takeaway for users is to move beyond basic prompts and start utilizing structured planning, research modes, and parallel agent execution to maximize productivity.