One agent with three lenses. Atlas plans in the Network conversation, works the library's own local tools first, and stages every write as a proposal you accept or reject. The archetypes below are its capability namespaces, not separate minds.
Names what a thing is. Tags new images against your existing vocabulary first and mints a new keyterm only when nothing fits; keeps meta tags, duplicates and the history honest.
Decides what belongs together. Finds, filters, sequences and sorts the canvas, and because its sort keys are the Archivist's own keyterms and palettes, the hand-off is invisible.
Files and ships. Keeps folders digital in Atlas, mirrors them to disk as real files under ~/Atlas Exports when you ask, and always asks where a thing should live. Originals are never touched.
There is exactly one agent. The archetypes are capability namespaces Atlas reasons through, not separate minds: nothing has to be explained twice and no archetype can disagree with another.
Archetypes coordinate through the same vocabulary, palettes and fingerprints, never through messages you can't see. The Curator sorts with the Archivist's keys; the Media Manager files what the Curator built.
Every archetype holds one of two levels today. ACTS may change what you see (the field, an arrangement) but never the library. PROPOSES may only stage changes for your accept.
Exactly one door writes agent intentions into the library: an accepted proposal. Atlas never creates, files, merges or deletes directly: you accept, or nothing happens.
Fingerprints, keyterms, palettes and similarity are free. Atlas reasons against local tools by default and spends model tokens only where language or vision genuinely helps.
Four doors (conversation, CTAs, “/” commands, contextual triggers) route into the same actions. Every tool call renders in the thread as it happens, and /history keeps the session’s record.