Why I switched my AI webmaster offGrok Bot email and back to AgentMail

Eight days ago I wrote that Grok Bot's built-in email beat AgentMail for my AI webmaster. Last night at midnight, I moved the setup back to AgentMail in 30 minutes.

After a full week of real client requests, the lesson is simple: an inbox designed for a personal assistant sending as you has different safety rules than an inbox built for an autonomous worker answering a client on its own. If every reply waits on your click, you have not automated the job.

How my AI webmaster works when email runs on its own

My AI webmaster, Sunny, runs a nonprofit client's website on Grok Bot, my cloud agent app. When I first automated myself out of the middleman role in late September, the goal was to take myself out of the daily relay loop so the client could work with Sunny directly.

Her workflow is straightforward:

• She emails a request in plain English. It might be a text update, a new video, or updated captions on a page.
• Sunny makes the change and gets a second review. A separate AI reviewer checks Sunny's code edit before anything goes back to the client.
• Sunny replies within minutes with a private preview link. Nothing touches the live site until the client looks at the preview and emails back "looks good".
• Every email copies me. That two-person rule keeps me in the loop without making me the courier for every change.

The entire service depends on one thing: email going in and out at any hour without waiting on a human to press a button.

Why I switched to Grok Bot's built-in inbox on October 9

When Sunny first launched, it ran on AgentMail, an email service built for AI agents. Sunny checked that inbox on a 15-minute timer, and every acknowledgment, preview link, and publish confirmation went out automatically without waiting on an approval click.

On October 9, Grok Bot started handing out free built-in email addresses. I switched Sunny over to the built-in inbox so incoming emails would wake the agent up right away instead of waiting on a 15-minute timer.

On October 11, after testing the new setup, I wrote: "My order now is the Grok Bot email first, and AgentMail for the rest."

Day one felt great because incoming mail arrived right away. Whenever the client or I sent an email, the trigger fired and woke Sunny up immediately.

Where the built-in inbox broke down in real client use

The catch showed up on the send side a few days later. Between October 12 and October 18, four problems turned an autonomous service back into a manual chore:

1. Background runs could read mail, and could not reply. When a client email triggered a background run, that run had no send tool. It could read the incoming message, and had to hand the reply off to the main chat.
2. Every reply sat as a draft waiting on my click. For about a week, outgoing emails could only go out as draft cards that I had to click Send on by hand. A system built to run on its own now needed me to press a button for every reply.
3. A real client sign-off stalled on a manual button. Last night at 9:22 PM, Sunny sent the client a preview of captions for a teacher intro video. She looked at the preview and emailed back, "Yes. Both look good. Great job!" Sunny caught her email right away and did the work, and still waited on my click to send the confirmation. The agent was doing the work. I was the router.
4. Turning off drafts traded draft cards for approval pop-ups. When I turned off drafts so Sunny could send directly, every outgoing email popped up a safety approval box because the platform treats every outgoing email as a high-stakes action. Even after I added our four addresses to the allow list and reworded the rule to cover every recipient on To and CC, the approval box still came up at 10:41 PM.

Worse, Sunny could not see that its own replies were blocked. At one point Sunny told me my test email had gone out fine while the approval box was still sitting on my screen waiting for my click. When I sent test emails at 9:54 PM and 10:41 PM asking for a reply, neither got an acknowledgment because both stalled on unseen approval pop-ups. From a client's side, a silent stall looks just like your webmaster ignoring you.

Moving back to AgentMail in 30 minutes at midnight

At 11:30 PM last night, we switched back.

Because I had never deleted our original AgentMail inbox when we tested Grok Bot's email on October 9, the rollback was fast:

• 11:30 PM: I reconnected Sunny to the original AgentMail account so it could see the webmaster inbox and the client's message history again.
• 11:59 PM: Sunny sent a test email from the AgentMail address. It went out immediately with no approval pop-up.
• 12:00 AM: We connected AgentMail's arrival alert so new messages wake Sunny up right away, replacing the old 15-minute timer with the same instant speed that made me try Grok Bot's inbox in the first place.
• 12:01 AM: I sent a one-word test email: "yoyo". The alert fired right away, and Sunny read it and replied on its own in 36 seconds with no human involved.
• 12:03 AM: With my OK, Sunny emailed the client (with me copied) asking her to send website requests to the original AgentMail address again.

Total time from decision to a working, autonomous round trip: about 30 minutes, after a week of stalled replies.

Three rules for giving an AI agent its own email address

If you are setting up an inbox for an AI agent in your business, here are the three rules I use now:

1. Separate assistant mail and worker mail. Approval pop-ups make sense when a personal assistant is drafting emails from your own address. When a dedicated agent is running a bounded workflow with a known client and you are copied on every thread, it needs an agent email service built to reply on its own, plus an arrival alert so it wakes up right away.
2. Test the round trip yourself. Never take an agent's word that a message went out. An agent will tell you its email went out fine while a safety pop-up is still blocking it on your screen. Send a "yoyo" test from your own phone or inbox and wait until the reply actually lands back in your inbox.
3. Keep your old setup alive for a week. Whenever you swap a working tool for a shiny new platform feature, leave the old inbox and credentials in place for at least a week of real client traffic. Keeping our old AgentMail inbox alive is the only reason rolling back a week of headaches took 30 minutes at midnight.

If my AI webmaster waits on my click for every email, I have not automated the work. Moving back to AgentMail let Sunny answer on its own again.

What is a tool upgrade in your business that you rolled back after a week of real use?