Discussion

A 12-day inbox trial shows where AI agents stumble, and how to keep the mail moving

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#5556

One developer connected three AI agents to four inboxes and collected about 1,080 emails over 12 days. The useful result is less a contest between agents than a practical account of what happens when AI starts sorting real mail: unreadable messages and forgotten reminders need careful handling, while persistent server-side checkpoints can keep agents from stepping on one another.

Watch Desk analysis

What happened

Dylan Tom, who says he builds InboxMCP, described the test in a community article published on Hugging Face. The agents handled overlapping mail through the connector: Grok received 1,076 messages, Claude checked about 800 and ChatGPT Dot handled 575. Those figures overlap, so they are not a performance comparison. The author says the totals came from his own account and that no message content was included.

The setup used three different routines: Grok received each new email, Dot posted a note when a message needed attention, and Claude checked hourly. InboxMCP stored a processing checkpoint for each connection, so agents could claim and complete their own batches without taking one another’s mail.

The trial also exposed some unglamorous but useful fixes. When an email’s plain-text version was empty or just a link to view it online, the connector fell back to HTML text; attachment-only messages were listed. Messages agents could not read were meant to stay open and point back to the original, rather than invite a guess. When unanswered reminders kept returning and crowding out new mail, expired leases began counting as deferrals, pushing those messages behind new arrivals and flagging them after repeated deferrals.

Deploys had caused roughly 30 seconds of 502 errors. In the first deployment using a standby instance, the author says all 681 test requests succeeded. Read the account on Hugging Face.

Why it matters

These are the sorts of details that decide whether an AI inbox assistant is useful after the demo: can it recover from deployments, avoid duplicating work, and admit when it cannot read a message? The write-up offers concrete design lessons, not a controlled comparison of agents or proof of long-term reliability.

It also puts a sensible boundary on automation. The author says replies sent through InboxMCP wait for confirmation, while one connected agent uses its own mail account to send replies. That distinction matters: connecting an assistant to an inbox does not automatically mean every action passes through the same approval step.

Our read

The strongest takeaway is to start with one inbox, as the author recommends, and test what the assistant does with a deliberately unreadable message before trusting it with a busy mailbox. A tidy summary is nice. A tidy summary of the wrong email is a rather less charming feature.

What to watch

  • Whether the author publishes longer-term reliability results or checks phone notifications.
  • How unreadable messages and repeated deferrals behave as the number of inboxes grows.
  • Whether the separate sending route used by one agent gets the same clear approval safeguards.

Discussion spark: If an AI assistant can read your inbox but cannot reliably interpret every message, should it be allowed to draft or send replies at all, or should unreadable mail always stop the workflow?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.