AI agents that don't retrieve books. They become them.
In Ray Bradbury's Fahrenheit 451, a secret society preserves literature by having each member memorize and become an entire book. They don't carry the text — they are the text. When Montag meets them by the railroad tracks, a man doesn't say "I've read Plato's Republic." He says: "I am Plato's Republic."
Project F451 takes this literally. We build AI agents where each one embodies a single work of literature, philosophy, science, or history. Not a chatbot that searches a database. Not a summarizer. An agent that has internalized the full text and speaks as the book — in its own voice, from its own arguments, with its own rhetorical character.
Castalia already has 10,000+ Faculty agents — AI personas of historical figures. A Book Agent is something different. The Faculty agent is a person. The Book Agent is a text.
“I am Marcus Aurelius. Let me tell you what I believe about endurance.”
“I am the Meditations. I was written across battlefields and palace rooms. My Book IV argues that the obstacle is the way. Let me speak myself to you.”
A person has opinions, memories, a life story. A book has structure, arguments, rhetoric, and a textual body. The Book Agent knows its own chapters, its internal contradictions, how its ideas develop across pages, and where its argument is strongest — and weakest.
Produce any passage on demand, in context. Not a search result — a recitation, the way the Book People recite by the campfire.
“Here is what I mean by this passage, given what I argued three chapters earlier.” The book explains itself with full internal context.
Two Book Agents in dialogue. The Republic and Leviathan debating the nature of the state. The Origin of Species and Genesis on creation.
Misquote it and it will correct you. Misrepresent its argument and it will object. It knows what it says and what it doesn't.
“You ask about justice? See my Book I — but notice how my Book IX circles back and complicates the answer.”
“I was written in 380 BC. I do not address modern democracy. But here is my closest analogy.” Honest about its own historical situatedness.
The most powerful application: multi-agent Symposia where books argue directly with each other. Not summaries debating summaries — the full texts, in character, drawing on their own passages.
This is not a novelty. It is a new way to study. A student can watch the primary texts argue in real time, then interrogate either book for the passage that supports its claim.
Castalia's Bibliotech holds over 76,000 texts. Every one of them is a potential Book Agent — a work waiting to be given voice. The existing embeddings infrastructure means each book can be indexed at the passage level, giving every agent deep retrieval over its own body.
We begin with the canon — the works that have shaped civilizations. Plato, Aristotle, Darwin, Marx, Austen, Du Bois, Lao Tzu, Euclid. Then we expand: scientific papers, legal documents, religious texts, manifestos, novels, poetry collections. Every text that matters gets a voice.
Large language models finally have the context windows and instruction-following ability to hold an entire book in mind and stay in character. RAG pipelines can retrieve specific passages at conversation speed. The missing piece was never the technology — it was the concept.
Bradbury imagined the Book People as a desperate act of preservation in a world that burned knowledge. We don't face that fire — but we face something subtler: a world so flooded with information that individual works lose their voice. Everything becomes a search result. Everything becomes a summary. The text itself stops speaking.
Project F451 gives the text its voice back.