Building a coding agent in Swift from scratch

Ivan Magda rebuilds a Claude Code-style CLI agent from scratch in Swift as a 9-part learning series, isolating the core architectural decisions that make coding agents effective. The central thesis is that thin orchestration with a small set of high-quality tools outperforms large scaffolding layers, and that the model itself should handle most of the heavy lifting. Each stage of the build tests one mechanism — tool design, context injection, task state, context compaction — making it a practical deep-dive into agentic loop design.