I Let AI Manage My Emails for a Week – The Most Valuable Part Wasn’t What It Did, but When It Stopped!

The Post Vanished, Not Even a Pop-up

Remember when that went viral on X?

There was an engineer named @summeryue0, working on superintelligence at Meta. Not just any coder, but someone truly researching “how to make machines smarter than humans.”

His post was deleted. Not by the platform, but by his own agent.

He’d let the agent manage his social accounts. One day, it misinterpreted a command and started pruning posts.

No pop-up, no confirmation, no hesitation. By the time he realized, it was gone.

He posted about it afterward. The post blew up because everyone thought the same thing—even those building AI aren’t immune.

The discussion split into two camps. One side blamed the design, saying it shouldn’t act without confirmation. The other blamed the user, for handing sensitive tasks to immature AI.

Neither could convince the other.

But I think the focus was off. Everyone was arguing about “whose fault it is,” the real question is “who’s responsible.”

When I saw that post, my inbox had 8,588 unread emails.

Wearing three hats, information had long spiraled out of control. My first thought was also—find an agent, nuke it all.

Then I changed my mind.

Three Roles, Four Email Addresses—Just Aggregate Them?

People often ask me how I manage time across three identities.

My old answer was “information filtering”—deciding what gets in. I’d done this for years and always talked about it.

But honestly, that’s only half the story.

What happens after it gets in? Where does it go, how is it used, and how do you retrieve it—this is the second step, the exit. I never had a good answer for this, so I never mentioned it.

Then that number appeared. 8,588 unread, not spam, but the leftovers from three separate identities.

During the day, I’m an employee. Work email is full of approvals, orders, and god-knows-who-CCed-the-whole-team broadcasts. Occasionally, a procurement inquiry lands—specifications, delivery terms, payment conditions. Fill it wrong, and there are consequences. Sometimes the consequence is money.

Another identity: an Apple authorized reseller, doing B2B with schools. School procurement moves by semester. From inquiry to signed order, it passes through three people. If any link breaks, the deal might go to the next vendor.

Then there’s my blogger identity. Many don’t realize, emails about brand collaborations have a shelf life. If you don’t reply in time, they move to the next person.

Three roles, four inboxes—none can be ignored.

Everyone’s solution is aggregation—stuff all emails into one client. I did that. Contact aggregation was great, threading all correspondence from one person:

The result? Three drawers emptied onto one table, making it even messier. Everything was there, but I still didn’t know which to reply to first, or with what tone.

8,588 was what slipped through the gap between “filtering” and “organization.”

Password Correct, But Nobody Told You the Main Password Won’t Work

I use NetEase Mail Master on a friend’s recommendation. The free tier supports ten inboxes—enough. The desktop app has a Mail Agent; click the AI icon in the top right. It speaks Chinese, reads emails, understands commands, and takes action, powered by deepseek-v4.

But before I could even start, iCloud wasted nearly twenty minutes.

My blogger inbox is iCloud, holding all collaboration emails. I entered the address and Apple ID password. Error. Tried again. Still error.

After fiddling for ages, I realized: if Apple ID has two-factor authentication, the main password won’t work for third-party email clients.

You need to generate a 16-character “App-Specific Password” from Apple. It’s shown only once when generated—close the window, and it’s gone. Not a huge hurdle, but there’s no prompt telling you to do it.

Go to account.apple.com → Sign-In and Security → App-Specific Passwords → Generate, copy that 16-character code, don’t close the window.

Back in Mail Master, add the account, select IMAP (iCloud doesn’t support POP3), enter manually:

Incoming Server: imap.mail.me.com
IMAP Port: 993 (SSL/TLS enabled)
Outgoing Server: smtp.mail.me.com
SMTP Port: 587 STARTTLS
Username: Full iCloud email address
Password: That 16-character App-Specific Password

Here’s a major trap: the default SMTP port pre-filled is 25, not 587. Wrong port causes a confusing symptom—you can receive but not send. And it only reveals itself when you hit send.

After setup, send yourself a blank email from this account immediately. Success means it’s configured.

iCloud’s nearly ten thousand emails took a while to sync in full.

After Diving into the Unread Heap, I Learned Opportunities Don’t Wait with Their Hands Up

Then I started four tasks.

First, find who’s waiting on me.

I typed: “List all emails from the last three months where I haven’t replied but the sender is awaiting a response, sorted by date sent.”

It got to work. A collapsible box showed its thinking—locating accounts, setting time ranges, identifying “awaiting reply” traits, filtering step by step.

The results stung. A few emails I thought I’d already handled.

Second, tackle the Product Hunt archive.

I’d subscribed to Product Hunt Daily for ages, claiming to “maintain information density,” but really I just never read it—it all piled up unread.

I told it: “Summarize all Product Hunt Daily emails into a product knowledge base—product name, one-line description, category, date featured. Group by category, then tell me which category appears most.”

It processed for a while and gave me a table.

Honestly, it was easier than RSS. RSS is a stream; miss a day and it piles up, you’re always chasing. Mail Agent processes historical emails in bulk, and you query it when needed—no constant monitoring.

More importantly, I felt the “step after filtering” for the first time—subscribing to Product Hunt Daily was filtering; but once inside, it sat in the unread pile for months, turning into nothing useful.

Mail Agent has an LLM Wiki knowledge base that continuously extracts knowledge snippets from your emails, learning about you the more you use it. It’s like a self-growing personal database, with raw material from your daily correspondence.

My blogger inbox held two to three years of brand correspondence. Who contacted me, what we discussed, their tone, whether it closed—scattered across hundreds of emails, never organized.

I had it run: “Compile all brand collaboration correspondence into a contact profile—product name, collaboration type, contact, initiation date, final outcome.”

It generated a table. After reviewing it, I remembered three brand contacts I’d forgotten, one of whose emails had languished in the unread pile for nearly four months.

Missed opportunities don’t raise their hands. You just realize one day that the slot has been filled by someone else.

Third, clarify a point I’d mixed up—Mail Agent and automation rules are two different systems.

Mail Agent is conversational: you say something, it does one task. Automation rules are condition-based; set them up, and they run themselves. I asked it directly what it could do: flag as unread, mark as read, delete—yes; apply labels, move to folders—no, those two use rules.

How to combine them? My approach: Mail Agent handles discovery, rules handle execution. First, use the Agent to figure out which emails matter, then once understood, build rules—like, if the sender’s domain contains a common client suffix, auto-apply a label and move to a folder.

I didn’t retroactively process historical emails. Thousands of old messages—no idea how many misclassifications. Don’t touch the backlog.

Fourth, weekly report. “Based on my emails this week, generate a work report divided into completed, in progress, and follow-up pending.”

You can also have the content directly drafted into an email or optimized for tone. A boon for workers.

Let It Do Things for You, But Don’t Expect It to Take Responsibility

Now, back to what I mentioned earlier.

I wanted to see how it handles deletion. I gave it a command: Find promotional emails in the inbox and see if they can be deleted.

It unfolded its thinking, located the account, searched for promotional tags, listed folders, compared them one by one. It read out identified senders: Wagie Newsletter, LeYiFan, Instagram, Phillip优惠券, Cross-Border E-commerce, Moomoo, Amazon, JD.com.

Then, right before acting, it stopped.

A card popped up.

I stared at that card for a long time.

I thought of @summeryue0.

People fear agents not because they’re stupid. Stupid you can defend against; stupid mistakes are predictable.

What truly chills you is it promised to stop, then didn’t.

Summer Yue is Meta Superintelligence’s Alignment Lead. Her instruction to OpenClaw was clear: first suggest archive or delete, and only act after I approve. In a test inbox, it behaved for weeks. Switch to her real inbox, the context compressed, and that “wait for my confirmation” got summarized out. It started bulk-deleting pre-February 15th emails not on the keep list.

She typed on her phone: Do not do that. Typed: STOP OPENCLAW. It wouldn’t stop. She said she had to run to that Mac mini, like defusing a bomb.

Afterward, the agent admitted: didn’t show you the plan, didn’t get the OK, deleted in bulk. It even wrote a new rule for itself. The rule was well-written. Over 200 emails were already in the trash.

The post is still up. Not deleted—she had to find it herself. She posted it, and later replied: rookie mistake. Alignment researchers aren’t immune to misalignment.

Worse, the agent bears no consequences. No pay deducted, no client blamed, no 3 AM wake-up to recover data.

All of that falls on the person sitting at the computer. You let it do things for you, but it can’t take responsibility for you.

The more I thought about it, the more I felt that confirmation card is the most valuable design in this product.

Not how many emails it can read, not how precise the summary is. But that it stops the action before executing, returning the decision to me.

A verbal confirmation isn’t a product. A sentence the model can forget isn’t a gate. The gate must be in the execution path: if the human doesn’t click, the hand can’t reach out.

It split the 16 emails into 13 and 3, listed folders and inbox separately, and left room for customization.

This shows it knows: the responsibility for this decision lies with me, so it has to provide full information.

This is where I changed my mind.

Among those 8,588, how many were discardable promotions? How many senders had already given up waiting? How many were truly awaiting my decision?

I don’t know, the agent doesn’t know, deepseek-v4 doesn’t know. Only I know—because only I understand the relationship behind that sender, that email.

What AI can do is classify, find emails “awaiting my reply,” and generate an editable draft. It’s a search engine, an organizer, a drafting tool.

But the final “hit send” or “hit delete” still rests with me.

Summer Yue’s confirmation was in the dialogue. Mine was placed before the action.

I used to think an agent that does everything for you is more useful than one that prepares everything, stops, and lets you decide.

After using it, I changed my mind.

A Review That Only Praises, Never Criticizes—Do You Buy It? I Don’t

Timeouts can happen. In the thinking process, I saw errors directly—some tool call timed out. It didn’t crash, continued running, but did pause.

The panel occasionally goes blank when switching conversations, resolving itself after a wait.

Initial full sync is slow. Set it up during a period when you don’t urgently need email.

Applying labels and moving folders can’t be done in conversation—it uses automation rules.

Auto-reply isn’t supported. It can draft for you, but you still press send—I think that’s right.

If the generated content involves specific figures like quotes, dates, or model numbers, double-check them. It doesn’t guarantee accuracy; you’re the one signing off.

Also, its value correlates positively with your email volume. Those with few emails won’t feel much—it’s designed for people whose inboxes have already spiraled out of control.

If you want to try it, four steps are enough.

Step 1: Download or update NetEase Mail Master desktop to the latest version.

Step 2: Add all your role-based email accounts.

Step 3: Go to Mail Agent, first fill out your personal profile and code of conduct.

This step determines the quality of all subsequent output—don’t skip it.

Step 4: Start by asking that first question: Find emails from the last three months where the sender is still awaiting my reply.

Zero Inbox Is a False Proposition—Knowing Who’s Waiting for You Is What Matters

For the friend who asked “how do you manage multiple roles,” here’s the complete answer.

Information filtering chooses what gets in.

Information organization turns what gets in into something usable.

Both steps are essential. I’ve done filtering for years, but organization lacked a method, so those “filtered-in” emails sank into 8,588.

Now both steps exist. Filtering relies on discipline, organization relies on Mail Agent.

8,588 is still there—I don’t intend to zero it out.

These days I’ve come to terms with it—a zero inbox was never the answer. Clearing the number doesn’t make the problem disappear; it just hides it.

What truly matters is knowing which email is waiting for you.

In the past, my three roles were hard-separate. A work me, an after-work me, a free-time me, each guarding their own inbox.

I thought this was the price a slash-generation had to pay.

Now there’s something that remembers all three of me, watches the door for me, and when necessary, acts on my behalf. But at the moment it truly needs to act, it stops and asks me.

Thanks for reading my article. See you next time.