Ian Whitlock has a proposal for reviving Donald Knuth's fifty-year-old idea that codebases should read like literature — and unlike every previous attempt, this one has a workable answer to the problem that always killed it.
Literate programming means code and prose interwoven: a human reads the file as a continuous explanation while a tool extracts the executable code into its target destinations. The extraction step is called tangling. The idea never caught on commercially. The reason was simple — whenever code changed, the prose had to change too. Nobody kept up. The explanatory layer decayed into lies.
Whitlock's fix is structural. Drop an AGENTS.md file into the project instructing Claude or Kimi to treat an Emacs Org Mode document as canonical source of truth. The agent handles tangling — extracting code blocks into their target files — and rewrites prose explanations whenever code changes. The maintenance burden that killed literate programming becomes the agent's job, not the engineer's.
He's using the pattern in two places today: test runbooks in Org Mode where prose-explained steps sit alongside executable code blocks with results captured inline — effectively a Jupyter notebook inside Emacs — and documentation for manual processes. He's also floating a hypothesis he hasn't yet validated at production scale: keeping intent-describing prose in the agent's context window may improve generated code quality. The intuition is that explicit stated intent gives the agent more signal than code structure alone.
The constraint he identifies is Org Mode's tight coupling to Emacs. A Markdown-based equivalent with comparable metadata support would make the pattern accessible to most working engineers. For now it's Emacs-native.
The timing is not incidental. Engineering roles are shifting from writing code to reviewing and directing agent output. If agents are already responsible for keeping prose and code synchronized, the codebase becomes a readable artifact — one a new team member or a future agent can actually follow — without requiring unusual discipline from anyone. That's a different proposition than anything literate programming's advocates were offering before.