Skip to content
Assistant & search

Decide what the assistant may do

A list of capabilities with switches — and an explanation for the ones that are locked.

Last updated 1 min readFor workspace admins

One page, every switch

The agent control page lists every action the assistant can take and every source it can read, each with a switch. Everything starts on, so a product update never quietly turns something on for you.

The setting belongs to the workspace. One workspace's posture never reaches another workspace's assistant.

Reading this page is as much a workspace admin's job as changing it. A plain member does not see it at all.

Switching something off closes three doors, not one

When you switch an action off, it first disappears from the list handed to the model — so it is never proposed.

Second, if a model proposes it anyway, it cannot be recorded.

And third — the one that really holds — a proposal banked earlier is re-checked at approval and will not execute. So switching something off also revokes the plans already waiting. Those proposals are not stranded: rejecting them is always allowed.

What each switch does when it is off

SwitchWhen it is off
The document toolNever called at all; the message answers like any other question
Semantic searchFalls back to keyword search; answers still carry their citations
One source typeVanishes from results without the index being rebuilt
One actionNever proposed, never banked, never executed

Locked rows explain themselves

Some rows have no switch: either they are invariants whose absence would make the product unsafe, or your plan does not carry that capability.

Either way the row states its reason. A capability that is absent has to say why, or the reader concludes something is broken.

Still stuck?

If the answer wasn't here, write to us. It is a real inbox and it gets answered.

Talk to us