Read path vs write path

The glass box shows you everything, and yet you can't hand-edit any of it. That sounds like a contradiction until you see the rule behind it: the glass box is the read path; chat is the write path. Everything is visible; nothing is edited in place. To change what Dustav knows or what's on your calendar, you ask it, and it makes the change itself.

Why not just let people edit directly?

Direct in-place editing is the obvious feature, and we rejected it on purpose, for two solid reasons:

• Security. The agent's context — the household facts it knows, the calendar it reads, the notes it drafts — is its highest-trust input; it shapes everything the agent does. A text box that writes straight into that context is a prompt-injection surface aimed at the most sensitive part of the system. Routing every change through Dustav, which edits coherently with its own tools, closes that door.
• Coherence. Hand-edited state goes incoherent fast. A human deletes a fact and leaves a dangling reference to it somewhere else; the agent, editing through its tools, keeps the record consistent. The agent is the right author of its own knowledge base because it understands how the pieces fit.

So the calendar, notes, and facts are read-only in the viewer, and the seam that could allow direct editing is kept switched off — held in reserve for a possible future power-user mode, not exposed today.

You change things by asking

"Ask Dustav to change it" is the tool for everyday edits — move the dentist appointment to Thursday, forget that we're vegetarian, that's not true anymore, tighten up that draft. You describe what you want in plain words; the agent finds the right event or fact or note and changes it, and you watch the change land in the glass box. There are no forms to fill and no fields to hunt for.

The escape hatch is deterministic

For the big, non-negotiable actions you don't want mediated by a model, there's a deterministic path that doesn't go through chat at all: delete your whole account. That's real, immediate, and account-level — the nuke, not a polite request.

The same rule covers the agent itself

This read/write split isn't only about your data — it's why there are no agent knobs. Dustav has a fixed name and manner; you don't reconfigure them from a settings screen. The one place with real direct controls is your own settings — your Anthropic key, billing, and the infrastructure dials like which model runs. Those are your stuff, not the agent's behavior, so a knob is exactly right there.

The throughline: see everything, change it by talking. You don't reconfigure the agent through a control panel — you tell it what you'd like to be different, and it does it.