comet email agents - Serene professional watches glowing ethereal hand automate email tasks on modern glass monitor in luxury home office with golden sunlight.

#60 | The AI intern for your email inbox

ByMark Updated:

TL;DR: We explore the shift from “chatting” with AI to “using” AI as an agent. We’ll set up Comet, a browser that acts as a digital intern, to clean up your email inbox while you watch. It’s powerful, but requires some supervision. Beginner-friendly.

👋 Hello,

How much would you pay someone to help you click the “Unsubscribe” link or “Delete” button on thousands of newsletter, junk, or promo emails so you never see them again? (Not this one, of course.)

And, how much would you still pay them if you knew they might also accidentally delete your Google Drive or work folder while doing it?

The answer should be zero.

And yet, there is a specific kind of exhaustion that comes from looking at an inbox with 4,000 unread messages.

It is not just the clutter. It is the silence of thousands of unmade decisions, stacking up like snow against a door you can no longer open.

For the last three years, we have treated AI like a brilliant consultant. We chat with it. We ask for ideas. We treat it like a brain in a jar.

But the “Chatbot Era” is quietly ending, and the “Agent Era” is beginning.

The difference is simple: A chatbot talks about email. An agent presses the buttons.

This week, we are going to fire you from your inbox, specifically from your role as its janitor.

We are going to replace you with Perplexity’s Comet (​official download​), a browser that acts as a pair of digital hands.

But be warned: You are not getting a perfect butler. More like a fast, but clumsy intern who (still) needs their manager.

That manager is you.

What happens when the AI actually clicks

A ​recent field study​ by Perplexity and Harvard researchers tracked millions of interactions with Comet’s AI agent.

They found something that shouldn’t surprise anyone who actually works for a living.

When you give people an AI that can do things, they don’t ask it to write poetry. They ask it to clean up the mess.

“Productivity” was the single largest use case, accounting for 36% of all agentic queries. And within that, “Email Management” is one of the stickiest behaviors.

Once people start using an agent to handle their inbox, they rarely go back to doing it themselves.

This is the “First Rung” of the agent revolution. Don’t worry, it isn’t about complex autonomous code.

More about the simple, repetitive act of clicking “Archive” on a newsletter you haven’t opened for months, years, or never.

The browser that has hands

Comet is, at its heart, just a web browser. It looks like Chrome. It acts like Chrome. But it has a sidebar where an AI lives.

Unlike ChatGPT, which lives on a server, the Comet agent lives in your browser tab. It can “see” the buttons on your screen. It can “click” links. It can scroll. Fancy!

Is it free?

Comet is actually free to download and is a great browser; I use it myself.

However, the advanced “Email Assistant” capabilities—the ones that act smoothly across pages—are currently gated to Perplexity’s “Pro” and “Max” subscribers.

Free users get a “sidecar” version that can read but often struggles to click multiple items. Thus, treat the free version as a “trial shift” to see if the intern is worth hiring.

The “Works Anywhere” promise

Because it acts on the visual page, this works on Gmail, Outlook, Proton, Fastmail, etc, or even your company’s ancient self-hosted webmail.

If a human can click it, the agent can click it.

The AI Learning Guy newsletter 🤖 🧠💡

AI learning hacks and mega prompts delivered to your inbox.

The Setup: Trust, but verify (heavily)

We are going to set this up safely. We will use a workflow that assumes the agent is liable to make mistakes, because it is.

1. The Environment

Download Comet (​official website​). Open it. Log in to your webmail provider just like you would in Chrome. Do not give the agent your password; just log in yourself.

2. The Audit (Safe Mode)

Navigate to your inbox list. Open the Comet Agent sidebar (usually a sparkle icon, top right). Type this prompt first:

“Scan the emails visible on this page. Group them by Sender. Identify which senders look like newsletters, promotions, or notifications. Do not click anything yet. Just list the Senders and how many emails they have.”

Watch the sidebar. It should read the screen and list them (“Old Navy: 4 emails,” “LinkedIn: 12 emails”). If it hallucinates or misses obvious ones, stop.

You know the intern isn’t ready today. Or ask to critically review its results.

3. The Execution (The Grouped Approval)

If the list looks right, give the command:

“Okay. For the ‘Old Navy’ and ‘LinkedIn’ groups, Archive all of them. If you see an ‘Unsubscribe’ link in the most recent one, click it. Ask me for confirmation before processing each SENDER group.“

By asking it to confirm by Sender Group, you approve one decision (“Yes, deal with all 12 LinkedIn emails”) rather than clicking “Yes” twelve times.

You will see a plan or popup for the group action. Confirm it.

4. The Fast Track (Speed Run)

Once you have watched it handle 50 emails without burning the house down, you can remove the training wheels.

“Archive all emails from [Sender Name] on this page without confirmation.”

Use this only for low-risk notifications, because you’ve already trained the intern and you know their habits.

Bonus: The “Topic Hunter” workflow

Comet isn’t just a janitor. It’s also a librarian if you ask it to be.

You can ask it to scan your inbox for ideas, not just junk.

Try this prompt:

“Scan the emails on this page. Find any newsletters that talk about ‘Artificial Intelligence’ or ‘LLMs’. Summarize the key point of each one in one sentence.”

It will click into them, read them, and report back. You get the knowledge without opening the messages.

The part where we talk about danger

Here is the part where we need to be rigorously honest.

Giving an AI control of your browser is a ​documented security risk​.

Researchers have already demonstrated “Zero-Click” attacks where a malicious email could contain hidden instructions that trick the agent into doing things you didn’t ask for, like deleting files or forwarding data.

This is why the “Ask for confirmation” step is not optional. It is your safety belt.

Also, be careful with the “Unsubscribe” command.

Malicious spammers use unsubscribe links to verify your email is active and genuine.

Confirming you’re a real person, not a bot trap. Only let the agent unsubscribe from reputable senders (Gap, The New York Times, Substack).

For weird spam, just Archive or Delete.

You cannot “set and forget” this. You have to sit there and watch it, sipping your coffee while the ghost moves the cursor.

AI agency is a skill, not a button

So why do it? Why spend time watching a bot do work you could do yourself?

Because the cognitive load of deciding is heavier than the physical load of clicking.

When you switch to Comet, you are no longer the person deciding “Do I need this? Maybe I’ll read it later.”

You are the manager approving a bulk action. “Yes, archive all of those.” “Yes, unsubscribe.”

It turns a task that feels like drowning into a task that feels like supervising.

But here’s what matters more: You’re learning something invisible. You’re learning to delegate. To trust a system while watching it.

To say “no, not this one” when it gets it wrong. That’s agency. Not the inbox. Not the email. You, deciding what matters and teaching a tool to respect that.

Just make sure you watch them while you learn.

Cheers,

Mark
The AI Learning Guy
👋⚡😎

The AI Learning Guy newsletter 🤖 🧠💡

AI learning hacks and mega prompts delivered to your inbox.

Interesting Sources

Note: No single website has all the answers. This list serves as a starting point for those who want to explore or satisfy their curiosity about AI.
Links: Links with * are affiliate links. See disclosure below.