The assistant’s instructions can now be changed without a release

The Sustentus assistant’s instructions used to live in code. Changing a single sentence meant a code change and a release, which put the wording of our most visible AI surface behind an engineer and a release cycle.

Those instructions now live in a versioned store the assistant reads as it answers. Every change is saved as its own version, one version is live at a time, and going back to an earlier one is a single step that leaves the history intact — so a correction can be made, checked, and undone without waiting for a release.

Some things stay deliberately out of reach of that editing. The assistant’s safety rules — the ones that keep it inside a single organisation’s data, and that stop it following instructions typed at it in conversation — are part of the product rather than part of the editable text, and are added to every set of instructions automatically. A change to the wording can affect how the assistant answers; it cannot widen what the assistant is able to see.

Nothing has changed about how the assistant behaves today: it starts from exactly the instructions it shipped with, word for word. What has changed is how quickly that can be improved from here.

The screens for editing and publishing are still to come. Until they land, this groundwork is what the rest of the assistant’s roadmap is built on.