Rubber Duck Thursday: Agentic Loops and Context Engineering
GitHub explores the context engineering step of an agentic loop, first testing context sharing in GitHub Copilot inside VS Code, then trying to replicate the same behavior using the Copilot SDK and LangChain.
Full summary based on transcript
What the stream covers
The presenter focuses on the “context engineering” part of an agentic loop:
- What context is provided to an AI coding assistant at each step
- How context is shared/maintained across steps in a loop
- How to reproduce similar context behavior across different tooling
Testing context sharing in GitHub Copilot in VS Code
GitHub tests how context sharing works when using Copilot in VS Code, looking at how the assistant can carry relevant information forward while working through a task.
Replicating the approach with Copilot SDK and LangChain
The stream then attempts to recreate the same context-sharing approach outside of VS Code:
- Using the Copilot SDK as the Copilot-side building block
- Using LangChain as the orchestration layer to implement an agentic loop
- Comparing the behavior and what needs to be explicitly provided as context
Key takeaway
The session is a practical exploration of how “agentic loops” depend on deliberate context engineering, and how context handling differs between an IDE-integrated Copilot experience and a custom loop built with the Copilot SDK plus LangChain.