Six Components That Make Coding Agents Work

Sebastian Raschka breaks down what actually runs under the hood of tools like Claude Code and Cursor into six architectural layers - from live repo context and prompt caching to subagent delegation. The piece lands on a point that keeps surfacing across the field: with base model capabilities converging, the harness around the LLM matters more than the model itself. Raschka backs this up with a minimal Python agent implementation, walking through context reduction, structured session memory, and tool orchestration as separate engineering problems rather than prompt tricks. For teams building their own coding workflows, this is closer to a reference architecture than a blog post.