An agent that can only modify a fixed set of rows has a hard time producing an open-ended system.
Cogweald does not completely predefine the cast of every world. The model can introduce a new character in a chapter, and the application stores that character with added_by = 'llm' so future chapters can use it.
The important part is not that AI can insert a row. It is provenance.
Every entity should have an answer to “where did this come from?” It may have been created by the creator, introduced by the LLM, or synchronized from an external tool. Once an agent can create data, provenance affects review, debugging, UI decisions, and future deletion strategies.
Open-ended generation needs two things:
- An extensible domain model
- A data model that tracks origin
Without provenance, an autonomous system eventually struggles to answer a basic question: why does this thing exist?
