Every conversation about giving an AI assistant write access ends in the same place: how do we know what it did?
There are two ways to answer. The first is reassurance, words on a security page. The second is evidence, a record you can read line by line. Only one of them survives contact with a real incident.
What a useful log contains
A timestamp, the instruction that triggered the action, the tool it touched, the exact change it made, and the outcome. Five fields, each boring, together a complete story of any action the assistant has ever taken.
The instruction field is the one teams undervalue. When something looks wrong, the first question is never what happened. It is why the assistant thought that was a good idea. An instruction-linked log answers that question in seconds instead of an afternoon.
Reversibility is the companion feature
A log tells you what changed. Rollback closes the loop. Anything the assistant sends, files, or edits should be able to describe its inverse, and every destructive action should pause for approval the first time it runs in a workspace.
None of this is glamorous work. It is the price of being allowed to operate in someone's real tools, and we think it should be paid up front rather than during the post-mortem.