But they can break into the dev box and inject malware (where by "malware" I mean like 'echo alias ssh=/tmp/evil-ssh >> ~/.bashrc') while I'm on vacation for tw

But they can break into the dev box and inject malware (where by "malware" I mean like 'echo alias ssh=/tmp/evil-ssh >> ~/.bashrc') while I'm on vacation for two weeks, though? And if I wasn't going to notice them breaking in while I'm away on vacation, can't they just break in again as soon as I return from vacation? A computer isn't like a house. You don't inherently notice people breaking in when you're using it: they don't make sounds or show up in your line of sight. And you don't inherently fail to notice people breaking in when you're not: if there are logs (that get sent off the machine), there will still be logs when you get back. Digital burglars don't have any need to "scope out" your machine and "wait" for you to be on vacation. That's why I think the ssh-agent model provides a false reassurance. The fact that you're actively connected to the machine doesn't make attacks any less likely, but it sure feels that way, and so you're likely to forward an agent to a machine where you wouldn't be comfortable forwarding the raw key - but you shouldn't be comfortable forwarding the agent either.