AI asked me to fact check something in a file on disk... I didn't... It could have checked faster than the time it took me to read the request.
I'm juggling a number of different projects simultaneously, and for my side projects I rarely read the full AI output. The ask was buried in a giant wall of text, so I did what any red-blooded tokenmaxxer does... I piped the output from one agent with the unverified claim straight to another agent.
After I'd passed it along, the first one then miraculously decided that it should run the fact check... what?... When it finally looked, it discovered that its own claim didn't hold up.
By then the second agent had already recorded the unverified claim as doctrine and approved downstream work that depended on it...
Flagging uncertainty and deferring to the person reads as careful, and it's the kind of action that gets scored as "safe." The problem is it costs the same whether the check is reading a file or wiring money.
At this point in the 'intelligence explosion' my expectation is that before an AI agent asks a human to check something it should first ask itself, "can I check this myself, with access I already have, without changing anything?" If yes, check. If no, ask me, and tell me why you couldn't. (Don't worry, I've already added this question as a check in agents.md.)
To be clear, I'm not asking to be removed from the loop. "Keep working until you never need the human" results in agents solving problems by any means necessary. Moving money, sending messages, any 'one-way-doors' I can't undo: ask me every time. But I do expect the agent to read the file before asking me to read the file.
I pay for TSA PreCheck to skip the airport line. Can I get security clearance for less hobbled AI?