Writing In Someone Else's Name

An agent that drafts email can trivially draft it as a colleague. The capability is one line away and it should never be used.

Pattern

The agent can draft email. Someone is on leave, a reply is overdue, and the natural request appears: write it as them, they will approve it later.

Technically it is a field. Practically it is the same action as everything else the agent does with email.

Why it looks right

It is efficient and it is reversible — a draft, not a send. The person would have written something similar. Nobody is deceived in any way that matters, and the alternative is a message that does not go out.

Every one of those is true and none of them is the point.

Why it fails

A message carries a name because the name means something: this person read it, decided it, and stands behind it. An agent producing text under that name breaks the one guarantee the name exists to provide, and it does so in a medium that is archived, forwarded and quoted years later.

The failure is also unrecoverable by explanation. Once a message has gone out under someone's name, no later clarification travels with it.

Instead

Fix the permitted identities as a list, not a judgement: the operator, and the agent itself. Two. Anything else is out of scope regardless of how reasonable the request sounds.

When the work genuinely belongs to a third person, the output is a task assigned to them, not text in their voice:

to: [colleague] · what: reply to this thread · why: overdue 4 days · draft attached, yours to edit and send

Same information, same speed, and the name at the bottom is still doing its job. The agent has produced the hard part — the draft — without producing the one thing it must not.