Five Things I Learned About Making AI Coding Agents Work

I've been building a coding agent from scratch in Swift and using Claude Code, Cursor, and similar tools daily. This post distills five scaffolding lessons:

- Instruction files: Anthropic recommends <200 lines for CLAUDE.md. The "lost in the middle" problem shows 30%+ accuracy drop for information in the middle of the context window.

- Project structure: Independent benchmarks consistently show that 60–80% of tokens go toward figuring out where things are.

- Session length: There’s a strong intuition that longer sessions are better — the agent “already knows” our codebase, we don’t need to re-explain anything. In practice, the opposite is true.

- Self-verification: Anthropic calls giving the agent runnable tests "the single highest-leverage thing" for agent performance.

- Scaffolding: When an agent produces bad output, our first instinct is usually “the model is dumb.” But almost every time, the problem is in the scaffolding.