Code Like a Surgeon (www.geoffreylitt.com)

The author proposes “code like a surgeon”: use AI agents to offload secondary, time-consuming engineering tasks so humans can spend 100% of their attention on the creative, high-leverage work they’re uniquely good at (e.g., UI prototyping). Practical examples of secondary tasks handed to agents include writing guides to relevant code areas, spiking big changes as sketches, fixing TypeScript errors or clear bugs, and generating documentation — often run asynchronously (overnight or while away). For primary work the author retains tight control and fast feedback (Cursor tab-complete favored), whereas for background chores they prefer long-running, less-visible agents like Claude Code or Codex CLI. This matters because AI now makes an old software model (Brooks/Harlan Mills’ “chief programmer” supported by helpers) economically and technically viable, reshaping team dynamics and status hierarchies: grunt work can be delegated to agents without burdening juniors. Key technical implications are clear — different autonomy levels need different tooling and UX (high-visibility, low-latency tools for core work vs. robust, long-running agent workflows for background tasks), and codebases must be structured to support automated agents. The author’s employer, Notion, both benefits internally from this setup and aims to productize the “surgeon” workflow for broader knowledge workers, emphasizing delegation of grunt work rather than replacement of core human expertise.