whats one action you never let an ai agent take without a human check
by•
whenever i give an agent write access i keep the riskiest action behind a separate check it cant talk its way past, because a prompt injected request looks exactly like a normal one. routing is plain code, the llm only writes the message. where do you draw the line, payments, sending emails, db writes?
5 views
Replies
For me the line is anything a real user receives. In my app's support panel the AI drafts a reply on request from a knowledge file, but a person edits it and presses Send, and only then does the mail go out. Credits can ride on a reply, and those are granted in the same transaction that records the reply, keyed by the reply's id, so a retried Send never grants twice. The model never touches the grant at all, it only writes words. On the coding side the same rule in a different shape: building, deploying functions and running scripts are the agent's to do, but destroying data and anything that reaches real users, like a store build, need a word from me first. Your "routing is plain code" is the key part, the check has to live somewhere the prompt cannot reach.