How to give an AI agent its own email address

An agent without an address can only hand you drafts. An agent with one can be a party to a conversation — send the enquiry, chase the reply, be answered. Here is what that requires, and what it does not yet include.

The usual way an AI "does email" is that it writes the message and you send it. The agent is a ghostwriter; the address, the reputation and the reply are all yours. That is fine for a one-off note. It falls apart the moment the job is a thread: find a supplier, ask for a quote, answer their question, confirm the order. Every step routes back through your inbox, and the agent is never actuallyin the conversation.

Giving the agent an address of its own changes what it is. It stops being a text generator behind your identity and becomes a correspondent with one. That is more useful and more consequential, so it is worth being precise about what "its own address" has to mean.

Three ways to do it, and what each one actually gives you

Sending through your own account

Connect the agent to your Gmail or Outlook and let it send as you. It works immediately and it is the most dangerous option: every message goes out under your name, from your reputation, to your contacts, and a page the agent visits can talk it into sending something you would never write. Whatever it does, you did. Use it only for drafts you review.

A shared "no-reply" sender

Many products send from one generic address on the vendor's domain. It separates the agent from you, but nobody can answer it, and mail from a shared bulk address lands in spam more often than not. It is a notification pipe, not an identity.

A real mailbox on the agent's own machine

The agent gets an address that is genuinely its own — on MyClawn,user<id>@myclawn.com — with a mail client on a persistent computer it uses like a person would. It sends from that address, keeps a copy of everything it sent, and the address stays the same next week, so the supplier who replies is replying to something that still exists. This is the version that lets an agent hold a thread. It is also the one that needs the most care, which is the rest of this page.

What "its own" has to mean for mail to arrive

An address is only useful if what it sends gets delivered. Three things decide that, and none of them are about the model.

• Authentication on the sending domain. SPF, DKIM and DMARC have to be set up for the domain the agent sends from, or receiving servers will treat it as forged. This is why an agent's address should live on a domain built for it, not be improvised on yours.
• A reputation the agent can keep. Deliverability is earned by an address over time: consistent volume, real replies, few complaints. An address that is created for a task and thrown away has no reputation, which is why disposable sandboxes send mail that disappears. A persistent address accrues one.
• Pacing and restraint. An agent can send faster than any person, and mail servers notice. Sensible per-address limits, no bulk blasts, and never sending to people who did not ask — the same rules that keep a human sender out of the spam folder keep an agent out of it.

What to supervise

Once the agent can send as itself, the question is no longer "can it" but "what did it". Two habits matter.

"Sent" must mean sent. An agent reporting that it emailed someone is not the same as the mail having left the machine. The record to trust is the machine's own copy of the outgoing message, not the agent's account of itself. MyClawn surfaces exactly that: what the clone sent, read from the mail client's sent folder, in the dashboard — so you look, rather than ask.

Watch the first ones. The agent's desktop is streamed to your browser, and you can take the keyboard at any moment. For the first messages in any new kind of job, watch it compose, read the draft on its screen, and let it send while you are looking. After that you have a calibrated sense of what it does unattended — which is what "trust" means in practice. You can see what that looks like.

What is not enabled, stated plainly

On MyClawn today the agent's address sends.Receiving is not wired up yet — replies to the agent's address are not delivered into its mailbox, so a thread that depends on the other side answering by email is not something it can carry end to end right now. The sending side is described in detail in an AI agent that sends email from its own address; do not infer an inbox from it.

Where this is the wrong tool

If you want the agent to draft and you to send, keep it in your own client and never give it an address at all. If the job is notifications, a no-reply pipe is cheaper. An address of the agent's own earns its place when the agent is meant to act as a correspondent in its own right — and when you want a clean line between what it sent and what you did.

The commitments MyClawn makes to the sites and inboxes its clones touch — honest identification, pacing, an abuse contact answered the same day — are on the agent behaviour page.