you are told to trust your AI. you are not told what trust is supposed to mean.
most of the time it means nothing. it means a logo, a confident tone, a promise on a marketing page. you hand over your inbox, your files, your calendar, and you hope. hope is not a security model.
here is the position this blog takes. an agent should do real work for you, and you should never have to wonder what it just did. it should act on its own, and you should be able to stop it with one move.
trust is not a feeling. it is a property you can check. you can read the client's source. you can watch what it sends on the network.
the veto is the mechanism, not the marketing. every outgoing message waits for your tap. one move stops it.
verification is the feature. trust is just what is left over after you check.
before you delegate, verify three things. which chats are excluded. what leaves your machine. who sees it. the AI provider you picked sees the prompt. mimicry's cloud does not.