I have been running Hermes Agent, the Nous Research one. Terminal, Telegram, a box that stays on. It writes its own skills after a messy task. That part is not vapor. You feel it the second time you ask for the same thing.
It is also not a company. It is a sharp tool with a learning loop, and it will happily wander if you do not point it. Fine for nights. Not how I would bill a client on Monday.
What I take from it: memory and skills beat a longer context window. The agent that remembers how you like invoices filed is more useful than the one that scores 2 points higher on a public bench.
I will keep breaking it. When something survives a week of real use, I will write that down here.