We recently shipped the Tiptap AI Toolkit (https://tiptap.dev/docs/content-ai/capabilities/ai-toolkit/o...), which lets developers embed LLM-powered editing into rich text editors.

Now we’re exploring the next step: making agents operate on documents without a browser involved.

The idea is a server-side system where agents can read, write, and transform documents in real time — essentially treating your document layer as a programmable data store that agents can interact with directly. Internally we’ve been joking about the name “real-time agentic document database”, because that’s the shape it’s starting to take.

I’m curious how others would think about this direction.

If you had a database where the fundamental unit wasn’t rows/objects, but documents with semantic structure that agents can mutate, what else could or should it do?

Where are the obvious footguns?

What boundaries would you want between agent autonomy and document integrity?

Would love to hear how you’d expand or constrain a system like this.