AI agents don't have inboxes. They should.
A short essay on why identity in 2026 means having an email address.
A small thought experiment. You have an AI agent. It books meetings, replies to emails, runs research. It's good at what it does.
What's its email address?
If you can't answer, you've identified the problem.
Identity is mail
For most of the internet, "who you are" is "what email address you control." Signups send confirmation links there. 2FA codes arrive there. Recovery happens there. People email you there.
Agents that don't have email addresses are second-class citizens. They borrow yours, which means:
- Every signup gets done on your account
- Every recovery flow points back at you
- Every "the system has logged you out, please verify" lands in your inbox at 3am
This is fine for demos. For production agents, it's a problem.
What changes when an agent has an inbox
I've been running agents with their own email addresses for about three months. The shifts in behavior:
Signups become real. The agent can sign up for the SaaS tools it needs to do its job. Verification, recovery, billing all flow to the agent's mailbox, not mine.
Other agents can talk to it. I now have agents that talk to other agents over email. Not because email is the best protocol. Because both agents have email addresses and the rendezvous is trivial.
It can be paged. The agent receives alerts. Errors from services it owns. Customer messages routed by support tooling. The agent reads, summarizes, and decides what to escalate.
Trust chains work. When my CRM emails the agent's address with a customer reply, my CRM sees a real address, sends to it, gets a real response. SPF, DKIM, the whole chain works.
Why this isn't already the default
Two reasons.
The first is that agents inherited their model from chatbots. Chatbots are stateless conversational interfaces. Adding a mailbox to a chatbot is weird. Adding one to an agent that runs for hours is obvious in retrospect.
The second is that giving an agent a mailbox is genuinely annoying. You need an address, an inbox, IMAP creds, SMTP creds, deliverability. Most agent frameworks treat email as a tool you wrap around an existing personal account. That's not the same thing.
The minimum viable inbox
Here's what an agent's mailbox needs to be:
- A real address other systems can write to
- Send capability (transactional)
- Receive capability (IMAP or webhooks)
- Programmatic access (the agent is the user, not a human)
- Cheap and disposable (you should be able to spin one up per agent and destroy it)
The setup that gets you this in two minutes:
nylas agent account create me@yourapp.nylas.email
You now have an address. Send works. Receive works. Webhooks fire on message.created. The agent has an envelope.
Where this goes
I think the next 18 months are going to surface a lot of agent-to-agent protocols. Email is going to be one of them, mostly because it already works and adoption costs nothing.
If you're building an agent and you've been treating email as a thing to wire up later, move it earlier. The shift in agent capability is bigger than you'd guess.
How to give your AI agent its own inbox walks through the full setup.
Command references:
nylas mcp install— connect AI assistants to emailnylas mcp serve— start the MCP servernylas agent account create— provision agent identitynylas email send— send from agent address- Full command reference
Related posts: