The mailbox used by Agent must first be turned into a work queue.

After I transferred a lot of newsletters to Agently Mail at night, my first impression was that it was actually quite straightforward: it was not “reading emails”, but sorting the inbox. The analysis that came back looked smooth, as if a bunch of messy inputs had been compressed into a few clues that could be processed further. This experience quickly raised a question - the real difficulty of email is never just understanding the content, but how to process the content after receiving it.

The mental model of a traditional mailbox is to receive, reply, and file. This model is sufficient for people, because people themselves are completing judgment, memory and tracking. After Agent comes in, this set of things is not enough. Agent is not here to “just read it”, it is more like a queue consumer: first determine whether the email is a subscription type, a notification type, or a follow-up type, and then decide whether to continue analyzing, generate a draft, transfer it to other tools, or directly throw it into the archive. As long as there is no such processing power, the so-called “smart mailbox” is just a more summarized inbox.

What I care more about is the traces left. The most annoying thing about emails like newsletters is not that you can’t finish the information, but that you don’t know which step has been processed after reading it. When people deal with emails, they will remember “I have read this one”, “I haven’t replied to that one yet” and “I will deal with that one later” in their mind. Agent does not have this kind of natural memory and must rely on status and logs to make up for it. There must be clear results as to which letter was skipped, which letter was summarized, which letter was converted to a draft, and which letter was asked for secondary confirmation. Without a result state, the Agent’s “help” will just end up becoming another more hidden layer of to-do piles.

This is also where this type of tool is most likely to go astray. Turning the mailbox into a chat interface may seem new, but if the bottom layer still remains at “summarizing a paragraph”, the value will be over. What is really useful is to turn the email into a transferable work object: the source is clear, the action is clear, the result is clear, and it can be played back if necessary. Only when this is achieved can the Agent begin to address the core issues of the mailbox, rather than just polishing the text on the periphery.

The boundaries are also clear. Interpersonal transactions, payment confirmations, contract back-and-forth, and replies with promises are not suitable for being handled directly by the Agent. High-stakes, low-format, responsible content is best left in human hands. What is more suitable for outsourcing to Agent is high-volume, low-risk, stable-format input: newsletters, system notifications, reminders, and routine summaries. What they have in common is not “worthless”, but “it needs to be dealt with every time, but it is not worth taking up people’s attention every time”.

The really interesting thing about something like Agently Mail is not making the mailbox more like a chat, but making the mailbox start to look like a replayable processing chain. What it solves first is not reading, but diversion, recycling and leaving traces. Once the mailbox changes from a communication tool to a work queue, the Agent has the opportunity to truly take over some of the daily chores.