While the idea is based on an initial concept I saw from someone else, and the expanded version is based on my own reasoning, the final writeup (this document) is AI-assisted. There will be revisions as the idea is refined further.
LLMs are surprisingly capable at completing tasks that fit neatly inside a single interaction. Give one a question, a document, or a small coding problem and it can often make impressive progress.
But stretch the task across several hours, many tool calls, multiple context windows, or a few separate sessions, and something strange begins to happen. The model may rediscover the same facts. It may forget why it made an earlier decision. A tentative assumption quietly becomes a permanent truth. A failed approach is attempted again. By the end, the conversation contains almost everything the model needs, but the useful state of the task is buried across hundreds of messages.