Working with the agent

The chat is not a help desk. The agent has the same access to your company's data that you have, and it does the work in place — creating records, starting processes, reading your documents.

Two modes

The switch above the message box has two settings, and they decide what the agent may touch.

Cowork — work with your data. Creating a customer, finding invoices, starting a process, answering a question from your documents. This is the default and where you will spend your time.

Extend — build and change apps. Adding a register your company needs, giving an existing app a new field. Here the agent changes the structure of your workspace rather than its contents.

Splitting them is deliberate: in Cowork the agent cannot restructure your workspace by accident, because it does not have the tools to. Switch when you want that, and not before.

It works as you

Everything the agent does, it does under your name and your permissions. A record you may not see, it cannot see; one you may not change, it cannot change.

That also means two people asking the same question can correctly get different answers. Yours is the one filtered to what you are allowed to know.

How far it goes alone

The second switch, under the message box, decides that: Ask & Plan plans without touching anything, Build works and asks before the consequential steps, Auto does not ask. See Ask & Plan, Build, Auto and when the agent asks first.

Writing a good request

Say what you want and why, and name the thing: "Add Meyer GmbH as a customer — contact Anna Weber, anna@meyer.test — and set up a project for the website relaunch."

Long is fine. Vague is expensive: "sort out the invoices" makes the agent guess, and it will guess differently than you would. If the request is long, dictate it rather than typing it — see dictating instead of typing.

The rest of this section