Agents should have room to exercise judgment.

The current operating model assumes that agents can:

  • split one raw file into several notes
  • infer likely destinations inside the garden
  • add wikilinks, tags, and cluster pages when the context supports it
  • move or rename assets when that clarifies the resulting structure
  • keep published prose free of workflow narration, provenance caveats, and other note-about-the-note commentary unless that commentary is itself part of the subject
  • avoid dangling wikilinks by linking only to existing notes or to notes created in the same change
  • keep the public atlas domain-first by preferring durable topical folders over workflow or placeholder categories
  • create a new subfolder only when a note cluster is dense enough to justify a reader-facing split

The review boundary is the commit, not a rigid up-front schema. That keeps the canvas flexible while still preserving editorial control.

This autonomy works best when the garden already values connected-notes, serendipity, and clear maps over brittle taxonomy.