How I handed my mail and calendar to an AI agent

What the agent does every morning

Every morning the agent reads the past day's mail. It filters out newsletters and notifications, and sorts the rest by project: films, agency, course, personal. For every important message it prepares a card: who is writing, what they want, what to reply. A draft reply is already inside.

I read the cards and decide. No message goes out without my go-ahead.

Once, in a single run, the agent sorted through 186 messages: it archived 154 and set aside 32 for me to answer.

Meetings go straight into the calendar

As soon as a message mentions a meeting, call or appointment with a specific time, the agent adds it to my Google Calendar and tells me: "added to calendar: what, when." If the event already exists, it does not create it a second time. It never adds people to an invite without my permission.

How to set it up

1. Install Claude Code or Codex.
2. Connect your Google mail and calendar in the program's settings. You enter your password yourself on Google's own page; the agent never sees it.
3. Give it a first task in plain language: "sort today's inbox, separate what matters from the newsletters, and suggest replies for what matters."
4. Write down the rules: what the agent does on its own, and what it only does with your permission. It remembers them from one conversation to the next.

The steps are the same in Claude Code and in Codex.

What went wrong

The agent made mistakes, and some of them cost me real stress. Here are three that made me change the rules.

Missed messages. At first the agent looked for new mail among senders and subjects it already knew. If someone wrote from a new address, or replied inside an old thread, the message slipped past it. Once, a reply to my own request, one with a two-hour window, sat unread for two days.

What I changed: the agent now scans the entire inbox, and it keeps watch on any thread where I asked for something, until a reply arrives.

False alarm. The agent told me three times about a conflict: a dentist's appointment supposedly clashing with a conference. It trusted the calendar entry. But the clinic had rescheduled the appointment long before, at my own request by email, and the calendar had simply not been updated. Over this "conflict" the agent went as far as clearing my meeting slots at the conference, and I had to put them back.

What I changed: before it says "conflict" or cancels anything, the agent now opens the latest messages on that arrangement, both mine and incoming.

Shifted date. Once, the agent logged an all-day event in a way that, because of the time zone, put it a full day earlier in the calendar.

What I changed: before any action involving dates, the agent now checks the date and time against the computer's own clock.

What not to give the agent

1. passwords and card details - I enter those myself;
2. permission to send without your go-ahead - the agent drafts the message, you decide;
3. money - the agent does not make payments or transfers.

Frequently asked questions

Is it safe to give the agent access to my mail?

You grant access yourself and can revoke it at any time in your Google account settings. You can also require your approval before it sends anything.

Does it read all of my mail?

Only what you tell it to. For me that is the inbox from the past day.

Do I need a developer?

No. Everything is set up in plain language.

How long does setup take?

On the intensive, we connect mail and calendar on the first day.