Try an assistant before you trust it

Test mode: a real run that touches nothing

You can now run an assistant in test mode. It thinks through the whole job for real — reads what it would read, reasons over live data — but every action that would change something is held back. Nothing sends, nothing posts, nothing gets written anywhere.

You see what it would have done, listed out, and a comparison against its last real run so you can tell what your edit actually altered.

This is the answer to the most reasonable hesitation about handing work to an assistant: you can watch it do the job once before it does the job.

Permissions, per assistant

Each assistant now carries its own rules about what it may use and what needs your say-so. A research assistant can be allowed to read widely and permitted to change nothing. A drafting assistant can be allowed to write, but only with approval.

It’s set per assistant, not once for the whole workspace, so a cautious setting on one doesn’t hobble the others.

Assistants draw charts now

If an assistant is tracking something over time, it can put a chart on its page rather than only a table — lines, bars, areas. Ask for the view you want and it builds it.