Long-form writing: an agent that plans, researches and writes a real document

Everything else an agent does fits in one reply. Long-form writing is where it produces something you hand to someone else: a regulatory filing, a market-entry report, a due-diligence memo, a tender response. Those documents have three properties that a chat answer cannot have — they are researched before they are written, they are structured, and they take longer than anyone will sit and watch.

So this is not "an export button on the chat". It is a production line beside the conversation: stateful, interruptible, resumable, with a canvas where you read and edit the result.

It plans before it writes

The first thing that happens is not writing. The agent recognises which document is being asked for — from the document types you define, each with its own outline, required information, knowledge bases and refusal boundary — and lays out the sections. Every section gets its own brief: what it must cover, roughly how long, and which facts it depends on.

That brief is not hidden. It is the text the model is given for that section, it is on screen, you can edit it, and you can rerun that one section after changing it. A document you cannot steer is a document you will rewrite by hand.

If nothing matches, the agent proposes an outline and asks you to confirm it. Either way the structure is pinned before a single word is written, and it stays pinned for the run — editing the type tomorrow does not renumber the document you started today.

Each section researches on its own

This is the part that separates a document from a long answer. Each section goes and gets its own facts: it searches, it opens and reads the page — not the search summary, which is a title and two lines while the figures live in the body — and it keeps every fact next to the source it came from. Your own knowledge bases are searched the same way, and take precedence.

Then the section is written from those facts only. Not from what the model remembers about the topic. That is why the interesting failure mode changes shape: instead of a confident invented figure, you get a section that says the figure could not be found — and the tool that checks this is code scanning the text for specifics that appear in no source, not the model grading its own work.

The research is bounded and budgeted per document, and it stops when it stops finding new things. You can watch which domains it read, per section, while it works.

It asks before it invents

A document usually needs a handful of facts only you have: the legal manufacturer, the target market, the filing deadline, the class of the device. Missing one of them does not degrade the document slightly — it invalidates it.

So the required information is part of the document type, and the agent checks it before starting: it asks for the one thing that is missing, says why it needs it, and holds. Nothing is generated, and nothing is billed, until the answer is in. If you already said it earlier in the conversation, it does not ask again.

You change a passage, not the document

Select a sentence, say "add the Brazil deadline here", and that passage is replaced. The rest of the section is byte-for-byte identical.

The mechanism matters more than it sounds: the selection range is computed by the system, and the model is only ever asked for the replacement text. A model asked to return a range does not return an error when it is unsure — it returns a plausible-looking range, and a sentence quietly disappears from your filing. So it never gets to choose one.

The same path handles a whole-section rewrite, a manual edit in the editor, and an instruction typed into the chat. And because the document sits beside the conversation rather than inside it, you can keep talking to the agent about the document while it is open — asking it to add a chart from your own data, pull in a file you just uploaded, or translate a copy into another language.

Nothing you had is ever lost

Every change to a section is kept: the version it was first written as, your manual edits, each requested rewrite, each restore — with what you said at the time, and which passage it touched. The comparison against the current text is computed line by line; a model is never asked to describe what changed, because a description that reads well and is wrong is exactly the failure this feature exists to prevent.

Restoring an old version adds a new version. History only ever grows.

What you end up with

A document in the canvas, exportable as PDF with a clickable table of contents and page numbers, with charts drawn only from numbers that appear in your material. A read-only copy can be translated into another language without touching the original. And when you share the conversation that produced it, you decide whether the document goes with it.

The honest limit: this is for documents that are worth a few minutes of machine work and a review by someone who knows the subject. It removes the blank page, the fact-gathering and the assembly — not the accountability.