The Agent Was Already 70% Built

PureTome’s AI used to write finished memoir prose, name the chapter it belonged in, and leave the pasting to you. This week that loop closed in production. It took one new tool, because most of the agent already existed.

The Clipboard Was Load-Bearing

The old interview feature did real work: it asked about your life, synthesized publishable prose from your answers, and generated a plan naming exactly where the material belonged in your book. Then it stopped. Getting that prose into the manuscript meant a side panel, a badge that literally said “Copy-Ready,” a clipboard, a walk over to the timeline, two creation modals, an editor, and a paste. Twelve steps across five screens, for one event, in a book that needs hundreds.

The system knew what to write and where to put it, and made the user perform the final step by hand. For a working writer that gap is friction. For the 70-year-old the product exists to serve, someone who will never learn a timeline UI, it is a wall. The core promise, you talk and the book gets built, was structurally false on that flow.

Most of the Agent Already Existed

The obvious fix was an agent, and the obvious fear was a rewrite. So before scoping anything I read the execution paths end to end, and the fear didn’t survive contact with the code.

A tool loop was already running in production, dispatching one read-only tool through a hardcoded if-statement. An authenticated write channel between the AI service and the API already existed. The rephrase feature had already proven the path where AI text becomes persisted manuscript prose. And every capability the agent would need, interviewing, critique, query letters, already lived as a discrete service waiting to be wrapped as a tool.

The Conductor, the LangGraph router that picks persona and phase for every conversation, stayed put. Nothing got rewritten. What was missing came down to wiring: a registry where the if-statement stood, write tools next to the read tool, and a surface where conversation is how you work instead of a side panel.

One Tool Before the Tool Zoo

The temptation was to ship fifteen tools and call it an agent platform. I shipped one.

One behavior proves or kills the thesis: you say something true about your life, and a real event appears in your book, placed by the model, no copy-paste, no screen-hopping. The first write tool appends to your memoir and does nothing else. It decides that the story about your father’s deli belongs in the childhood chapter, or proposes a new chapter when none fits. The kill criterion was written down before the code: wrong placements, or any feeling of danger, and we fall back to a plain “Add to memoir” button. Agents earn scope.

Constrained by Design

An AI writing into someone’s manuscript should scare you. Every constraint below exists because I assumed the agent would eventually get something wrong.

• Append-only. The agent can create. Overwrite and delete are absent from its vocabulary.
• Preview-and-confirm. It proposes; a card shows the chapter, the title, the prose; you approve, redirect, or dismiss. Nothing touches the book without your click.
• No new trust surface. Writes execute through the same authenticated calls your own clicks use. The agent got a voice, and deliberately did not get keys.
• Visible and metered. Every action lands live on the timeline, and every operation is billed like any other, so a chatty agent can’t drain your credits.

Live Now, Polish Next

As of today this is production: conversation is the primary surface and the agent mode is on. The blank-page fix shipped ahead of it, retiring the wizard that demanded a theme statement before you had written a single memory; the system now drafts your vision from the conversation and asks you to confirm it later.

I’m saying “live” and nothing shinier. A scoped debt list sits next to this launch: an expired session can still eat a typed message, the timeline loads slower than it should, some copy still describes the product this one replaced. That polish pass starts now. Shipping the loop first was deliberate. I needed the production answer to whether talk-to-book feels magical or frightening before spending sessions on fit and finish.

The Loop Is Closed

The “Copy-Ready” badge is gone. It was the most honest line of copy in the product: a label documenting the one step the system refused to perform. You talk; a chapter proposes itself; you say yes; the book grows.

The rest of the roadmap is refinement of a loop that now works end to end. That work starts tomorrow.