Inside The Agent Loop with Pierce Boggan

A deep-dive podcast conversation between VS Code team members explaining how the GitHub Copilot agent loop works internally. Topics covered include how system prompts are dynamically built per model, how tools are selected by the model, how sub-agents work (including the explore sub-agent that uses Haiku instead of Opus for faster context gathering), how VS Code optimizes agent trajectories using their own SWE-bench equivalent, and the tradeoffs between putting context in prompts versus using tools. The discussion also covers how users can customize or disable sub-agent behavior, inspect agent debug logs, and build their own custom agent modes in VS Code.