Use imbox

Agents in the app

Let an agent read new mail and write replies or organise threads, with limits the code enforces.

imbox docs

An agent works a mailbox for you. You choose the mailboxes it watches and write its instructions. For each new email that passes its rules, it reads the thread and follows them, either writing a reply or organizing the thread. Set them up under Settings → Agents, after you add an AI key. They use your AI provider.

This is different from Agent mail, which lets an outside program use a mailbox through an API key.

How an agent works

  • About once a minute, each agent that is turned on looks at mail that arrived since it was turned on. It never answers the backlog.
  • For each email it decides what to do from your instructions. It can read the thread, list labels, organize the thread (Reply Later, Set Aside, label, mark read, move to Feed) and write one reply.
  • Every decision is recorded, including the decision to leave a message alone, with the reason. Open Activity on an agent to see them. New drafts, sent replies and problems also show in the notification bell.

What happens to a reply

ModeResult
Write a draft for me to review (the default)The reply is saved in Drafts and announced in your notifications. Nothing is sent.
Send replies by itselfThe reply is sent from the mailbox as soon as it is written, marked as an automatic reply so other autoresponders leave it alone.

Start with drafts. Switch to automatic once you have read enough of its drafts to trust it.

Which mail it handles

An agent skips a message, and records why, when it is:

  • mail you wrote, or from your own other mailboxes
  • a newsletter or bulk mail, a no-reply address, an automatic reply or a delivery failure
  • older than a day
  • from a sender you have not let in (a setting, on by default)
  • from outside its list, if you gave it one of addresses or domains
  • in a thread it has already answered twice
  • over its daily limit (default 20, at most 200). Extra mail waits for the next day.

What it cannot do

The limits are enforced in code, not only in the instructions you write:

  • It replies only to the person who wrote, in the thread it was given, once per email. It cannot forward, add recipients, start a new message or reach other threads.
  • It cannot delete mail, trash a thread, screen senders in or out, search your mail or read memory.
  • In automatic mode the runner sends the draft. The model cannot send anything else.
  • A mailbox lent to an outside agent is never given to an agent here.

Write instructions that say what the agent is for, what it may answer, what it must never do, and when to leave a thread for you. Tell it to put anything it is unsure of in Reply Later.

Next

  • AI assistant: the key, the chat, and what is sent to your provider.
  • Agent mail: a mailbox for programs outside imbox.