How to Manage Multiple Gmail Accounts With One AI Agent

Work, the side business, and the account you regret. One agent that knows which mailbox it is acting in before it does anything — with a separate set of rules per inbox, because one rule set across several mailboxes is how a client invoice gets archived.

01The Problem

Four rules, one inbox

The work inbox needs its customers answered and its recruiters ignored. The personal inbox needs the exact opposite. Treated as one mailbox, every automation built for the work inbox eventually acts on the personal one.

One rule set across several mailboxes fails quietly

Nothing announces it. The personal invoice gets archived because the work rule set treats receipts as noise. A newsletter subscription gets auto-replied to because the work rule set drafts replies. Each individual action looked correct against the rules it was following — the bug is that it was following the wrong rules, and the symptom surfaces weeks later as something inexplicable rather than as an automation failure.

The context that makes one mailbox work is absent from the other

Your work inbox works because you know that the Tuesday thread is a live deal and the Friday thread is a supplier chasing an invoice. None of that is in the mail, and all of it is true of the other three accounts too. So a rule that generalises across accounts has to be a blunt one, because the specifics that make a rule good on one mailbox do not transfer to the next.

The account you least maintain is the one with the mess

Everyone has an address they signed up for and stopped looking at. It is usually the least business-critical and the most embarrassing to open, which is precisely why it stays open. A per-account scheduled run is the only version of this that gets done, because it needs no decision — it needs the rule set you already wrote for that mailbox, applied on a timer.

02The How-To

The multi-account prompt, step by step

Copy it once per mailbox, paste it into Zaira, and each one carries its own rules. The agent states which account it is acting in before every action, so the thing that makes this safe is the same thing that makes it slow: it names the mailbox every time.

Triage this mailbox

Step 0: State which mailbox this is

This prompt runs against ONE mailbox and must never act on another. Which account: [the exact address, e.g. work@mycompany.com]. What this mailbox is for: [e.g. client work for a 12-person SaaS]. What I am: [role]. Senders I always answer myself: [names or domains]. Senders that can never reach me here: [newsletters, cold outreach, recruiters]. Recurring senders worth keeping but not reading daily: [status pages, receipts, GitHub]. Timezone and working hours: [e.g. Europe/Lisbon, 09:00-18:00]. Tone for anything you draft: [e.g. short, no exclamation marks, sign as Alex]. What must never happen on this mailbox: [e.g. no auto-replies at all, never delete anything from legal]. Before any action, tell me which account you are about to act on. If the account you are connected to is not the one named above, stop and say so.

Step 1: Pull the backlog

Search this mailbox for the last 14 days using Gmail search operators, in several passes rather than one: unread, unread and older than 3 days, threads with more than one message from the same person, and mail from my priority senders. Deduplicate across passes. Report the count before acting.

Step 2: Classify every thread into exactly one bucket

- NEEDS ME: blocked on a decision or action only I can take - DRAFT IT: needs a reply, and I can give you the substance to write one - FILE IT: genuinely worth keeping, but not this week - CLEAR IT: no future value — receipts, expired offers, no-reply notifications, duplicates Show the counts per bucket before you act. If NEEDS ME is more than 15, say so and name the senders driving it.

Step 3: Draft the replies

For each DRAFT IT thread, read the full thread with get_thread, then create a draft in my voice using the tone rules above. Quote the specific detail their message contains that the reply answers, so I can see it responds to the right thing. Keep each draft under 120 words. Do not send anything. Add one line at the top of each draft saying what it answers and what I need to confirm.

Step 4: File and clear

Create labels for this mailbox's buckets if they do not exist, namespaced so they cannot collide across accounts. Apply them with batch_modify_labels — never re-read a thread to check its state. Move CLEAR IT to trash and FILE IT out of the inbox into its label. Leave NEEDS ME and DRAFT IT untouched.

Step 5: Report

Give me: the account name, the counts per bucket, the 10 threads that most need a decision with sender, subject, one line on why, and the date it arrived, every draft you wrote with a link, and anything you were unsure enough to leave alone. Then tell me exactly what you changed and what you did not touch. Stop.
03Why People're Using

What this does once it's running

You stop carrying three sets of mental rules. The side business gets the same daily triage as the day job without the work rules leaking into it, and the personal inbox finally gets cleaned up on the same schedule as everything else.

It names the mailbox before it acts

The first thing the agent says is which account it is about to touch, and it stops if that account is not the one this prompt was written for. Naming the account on every run is what makes running three of these concurrently safe rather than merely usually fine.

Rules cannot bleed between inboxes

The sender lists, the tone, the buckets and the never-do list are all per mailbox, and the labels are namespaced. The archive-the-receipts rule that is correct for a side business never runs against a personal inbox, because that prompt does not contain it.

The neglected account gets maintained too

Four accounts means four scheduled runs on the same pattern, including the one you never open. It needs no decision each time — the rules are already written — which is the only reason it happens at all.

04FAQ

Frequently asked questions

The questions people ask before running one agent across several mailboxes.

Not safely, and this is the most common mistake here. A prompt tuned for a work inbox will archive receipts, draft replies to cold outreach, and clear mail you needed, and it will do all of that on your personal account without any sign that it had been pointed the wrong way. The safe version names the exact address in Step 0, tells the agent to report which account it is about to act on, and has it stop if that account is not the one expected. The rules themselves should still differ per mailbox.

No. Each mailbox authorises as its own connection, so granting access to the side business account does not change what the agent can reach on the work account, and revoking one leaves the others in place. That separation is what makes several accounts safe to run at once — but it is the connection that is separated, not the rules, which is why the per-mailbox prompt still matters.

The names can match; the contents should not. The buckets are just NEEDS ME, DRAFT IT, FILE IT and CLEAR IT, but what lands in each is entirely different between a work inbox and a personal one — and that difference is carried by the sender lists and the never-do list in Step 0. Give the personal mailbox a much longer CLEAR IT threshold, because almost nothing there is genuinely urgent.

More than the accounts you use carefully, which is counterintuitive. The mailbox you never open is the one that has been accumulating for years, and a per-account scheduled run is the only thing that will ever clear it because it requires no decision from you. Run it on a longer window and a longer clear threshold than you would for a working inbox, and expect the first run to report a lot.

No, and it should not try to. Inferring which mailbox a rule belongs to from the content of a message is exactly the wrong design — that inference is what produces a personal invoice being archived by work rules. The separation has to be explicit, per prompt, which is less clever and is the reason it is dependable.

Whatever the connections support, and the practical limit is your attention rather than the tooling. Four accounts with a clean rule set each is a morning of setup and then nothing to think about. Beyond that the rules start overlapping in ways that are hard to hold in your head, which is the point at which a shared inbox with routing is the better shape.