The Chat Window Grew Hands: Grokbot, Muse, and Hermes

Here's the scene. A twelve-person shop. The office manager used to paste a messy paragraph into a free chatbot to clean it up. This month she installed something new because the ad said it would “handle follow-ups.” She signed in with her work mailbox. By Thursday it had sent three emails she did not write, from her name, to people who pay you.

Nobody broke in. She thought she was still chatting. She had hired a helper.

That is the rise of agents. Not a smarter essay. Hands.

Chat waits. An agent takes a job.

A chatbot sits there until you ask. An agent is given work: read the inbox, draft the reply, file the PDF, look something up, click the button. The window looks the same. The permissions are not.

You will hear names. Grokbot is the Grok-shaped one people meet on their phone and on X. Hermes is an open agent from Nous Research that runs on a computer you control, with tools and a memory that lasts past one chat. Muse is in the same pile of named helpers being sold as “it will just take care of that.” The brand on the icon will change. The pattern will not: talk to it, then it goes and does.

I am not going to recap feature lists. Those rot in a month. The question that does not rot is: what did you plug it into?

Why it looks like nothing happened

Connecting mail or files is one “Continue with Google” screen. It feels like logging in, not like handing a stranger the keys. Sending a message feels like the app being useful. There is no alarm that says “this left the building.”

Same force as staff pasting client notes into a chatbot on a deadline, which we already wrote about in Nobody Meant To Leak It. The new part is the send button. Paste is a leak. Send is a letter on your letterhead.

What it actually costs

Customer mail sitting with a vendor you did not pick. A wrong answer that already went out. A calendar invite from “you” that you never saw. If a client asks where their data went this week, “in the AI” is not an answer you can stand on.

You do not fix that by training a custom model. Training copies last quarter into a brain you then have to babysit. A helper that can look at this morning's records is the job. The records stay yours.

Treat it like a new hire

Bans fail. People will install the thing that saves Thursday. Give them a sanctioned helper, write two paragraphs of rules, and teach one line: if you would not email it to a stranger, do not paste it, and do not let the helper send it unsupervised.

If it can act, give it the same hygiene as a person:

• A login you can turn off without them in the room
• Access to the one system it needs, not everything
• You read what it sends until you trust it, the same way you would with a new assistant

That is not theater. That is how you fire it on Friday if Thursday went sideways.

A five-minute check

1. Has anyone connected an AI app to work email, calendar, or files?
2. Is that login the company's, or a personal account?
3. Can you disconnect it without that person present?
4. Has it sent or filed anything nobody reviewed?
5. If a client asked where their data went this week, could you answer in one sentence?

If you cannot answer those, you do not have an AI strategy. You have an unpaid intern with a keycard.