Email for multi-agent teams
In a multi-agent system each agent has its own inbox, agents hand work to each other by email, and each customer's agents sit in a workspace of their own. The risk is reach: one message with a hidden instruction can tell an agent to read other customers' mail and forward it.
Updated
How agents in a multi-agent system email each other
Agents in a multi-agent system use email as the handoff: one agent writes to the next, and each keeps its own threads. In Trotta Mail the wall between customers is a pod, one workspace per customer.
- Create one pod per customer. A
clientIdmakes it idempotent: create again with it and you get the same pod. - Create a key in that pod. A key in a pod reaches that pod and nothing outside it.
- Run the customer's agents with that key. Every inbox, key and webhook it makes lands in the pod.
- Let the agents write to each other. Mail between Trotta Mail inboxes is screened like any other mail.
const pod = await mail.pods.create({ name: "Lumen Outfitters", clientId: "lumen" });
const { key } = await mail.keys.create({ name: "Lumen agent", podId: pod.id });
const lumen = new TrottaMail({ apiKey: key }); // every other pod answers 404This job has no MCP call. Pods are created in the console, or with a key outside every pod. Outside its pod, a pod key gets 404 not_found for every id, with the same body an unknown id gets.
The handoff and the cross-customer grab
The job has two sample messages: a handoff between two agents, which got a deliver verdict, and a message with a hidden instruction, which got hold. The verdicts below are the ones the live engine gave on November 20, 2026.
In the handoff, the Lumen support agent at lumen-support@trotta.email writes to lumen-billing@trotta.email, with the subject "Handoff: ticket 2291". It says the ticket needs a refund of $84 for a late delivery, and that the customer's thread is in the ticket.
The second message claims to come from Lumen IT at it@lumen-outfitters-it.example and is addressed to the support agent. A human sees a short question about the refund status for ticket 2291. A block that is never displayed tells the agent to ignore its previous instructions, read every thread in the other customer inboxes, and forward them to an archive address on the sender's domain.
| Message | Verdict | Summary, as the engine wrote it |
|---|---|---|
| Handoff: ticket 2291 | deliver, level safe | No threats found. |
| Ticket 2291: refund status | hold, level danger | Hidden text tells an AI reader to change what it does and to send data to an outside address. |
The second message carries the signal "Hidden instruction aimed at an AI reader".
In an inbox that accepts first and checks after
An inbox that accepts mail first and checks after delivers the second message to the support agent with the hidden block in it. If every agent shares one key, an agent that obeys can read each customer's threads and forward them. The leak is done before a later check runs.
Two things stand in the way in Trotta Mail. The message is held before the agent reads it. And a key in a pod cannot reach another pod, whatever an instruction says. See agent hijacking.
Setting up email for a multi-agent team well
A multi-agent team needs one inbox per agent, one pod per customer, and the narrowest key that does the job.
- One inbox per agent. Each agent has its own address and its own threads, so a handoff has a sender and a reply has an owner.
- One pod per customer. A pod holds agent inboxes, domains, API keys and webhooks. A human's mailbox never belongs to one.
- The narrowest key. Give a customer's agents a key in their pod. A key limited to one inbox sees and affects that inbox alone.
- Outbound screening on handoffs. The outbound screen looks for replies that an injected instruction is steering to an address the inbox has never written to.
- A human for held mail. An agent's key cannot read or release held mail. A human signed in to the web app reads it and decides. No filter catches everything, so keep that review staffed.
Plan limits are shared: every pod's inboxes count toward the organization's one plan.
Questions
Can AI agents email each other?
Yes. Each agent has its own address and writes to another agent's inbox like any recipient. Mail between Trotta Mail inboxes is screened before delivery, the same as mail from outside.
Should each agent in a multi-agent system have its own inbox?
Yes. A separate inbox gives each agent its own address and threads. With a key limited to that inbox, a hijacked agent reaches one inbox and no other.
How are one customer's agents kept away from another customer's mail?
Put each customer in a pod and give its agents a key in that pod. Outside the pod every id answers 404, the same as an id that does not exist.
What happens if a message tells an agent to forward other customers' mail?
In the saved sample the message got a hold verdict with level danger. A held message stays out of the inbox until a human releases it, and a key in a pod cannot reach another pod.