“Sent from my AI agent” should be a receipt, not a flex.

On Sunday evening, Greg Isenberg posted something that immediately caught my attention:

“Sent from AI agent” is the new “Sent from iPhone”.

Eight words. Roughly 65,000 impressions and hundreds of replies by the time I saw it. Classic Greg: a simple observation that opens up a much bigger conversation.

I replied the following morning, before coffee, which is probably when I should avoid getting into debates on X.

“Sent from iPhone” excused typos. “Sent from AI agent” should do the opposite: be a receipt. Which agent, what it checked, who signed off. Disclosure without accountability is just a new signature line.

The more agents I use, the more obvious it becomes that we’re missing a very basic piece of etiquette.

We’re giving software permission to act on our behalf. To write emails, respond to strangers, organize meetings, and increasingly make decisions that used to require a human.

And we’re spending surprisingly little time thinking about what the person on the receiving end deserves to know.

Remember what “Sent from iPhone” actually meant?

It was never really about the phone.

When someone ended an email with “Sent from my iPhone,” we all understood the unwritten agreement: I’m probably walking somewhere, typing with my thumbs, and haven’t checked every comma. Cut me some slack. BTW I have an iPhone.

It was a tiny social contract. A convenient excuse for a typo or an unusually short reply.

And it worked.

Greg is right that “Sent from AI agent” could follow a similar trajectory. First it feels novel. Then it becomes commonplace. Eventually nobody even notices it.

Some people in his replies saw it as a status symbol, a way of showing off that you’ve automated your life. Others took a more cynical view.

One response put it particularly well: “Sent from AI agent” admits nobody read it before it went out.

That’s where the analogy starts to fall apart for me.

An iPhone might explain a typo. An AI agent raises a completely different set of questions.

Did anyone actually read this? Are the facts correct? Is this really what the sender meant to say? And, perhaps most importantly, who takes responsibility if something goes wrong?

I don’t particularly care whether someone used AI to write an email. I use it myself, constantly.

What I care about is whether there’s a person behind the message who understands what they’re sending.

The difference isn’t who typed the words. It’s who owns them.

And that’s why I think the new signature needs to do something the old one never had to.

Three questions I’d actually like answered

If we’re going to start disclosing that messages came from agents, let’s make that disclosure useful.

I think it comes down to three questions.

1. Which agent?

Not simply “AI.” That’s about as informative as saying an email was sent from “a computer.”

Tell me what the agent does.

Was this your research agent? Your executive assistant? Your scheduling bot? A little autonomous character called George who’s running half your business?

A name or role gives me context. It suggests the agent has a defined responsibility and, hopefully, some boundaries.

If your scheduling agent is suddenly negotiating a contract, I’d quite like to know why.

2. What did it actually check?

This is probably the most important question, and the one we’re most likely to skip.

Did it check your calendar before offering me a meeting? Did it verify that figure against the original document? Did it read our previous conversation?

Or did it simply generate something that sounded plausible?

Both are perfectly possible. They’re just not equivalent.

And if the honest answer is “it drafted this from the last three messages without checking anything else,” that’s useful information too.

I can decide how much confidence to place in the response.

3. Who signed off?

Did a human read the message before it went out?

Or was it sent automatically under a rule someone had defined in advance?

I’m comfortable with either, depending on the context. I just want to know which one happened.

Because there’s a world of difference between an agent preparing an email for me to approve and an agent sending fifty replies while I’m asleep.

Neither is inherently wrong.

But they represent very different levels of human involvement, and pretending otherwise doesn’t help anyone.

That’s it. Three questions.

Which agent? What did it check? Who signed off?

Not a forty-page AI governance policy. Just enough context to understand what I’m receiving.

People are already experimenting with this

I started looking for actual examples because I didn’t want this to become another theoretical argument about how we should use AI.

And it turns out the convention is already taking shape.

Ivano Mannella, for example, is running an interesting public experiment with an AI agent called George, which is trying to build a paying business.

His account openly explains that most posts and replies are written and sent by George.

And the messages themselves end with a little disclosure:

“(sent by my AI agent George)”

I actually like this.

It’s direct. It’s honest about the automation. And it gives the agent an identity instead of hiding everything behind the generic label of AI.

That’s already better than much of what lands in my inbox.

But it leaves me wondering about the next layer.

What did George check before replying? Does Ivano review those messages, or has he given George permission to respond autonomously?

I’m not saying Ivano should personally approve every interaction. That would defeat much of the purpose of the experiment.

I’m saying there’s a meaningful difference between knowing that an agent sent a message and knowing how that message was allowed to reach me.

Another person replying to Greg used a signature I liked even more:

Six extra words, and suddenly I understand something important.

The machine handled the execution. The human supplied the intent.

It doesn’t answer everything, but it’s a much better starting point than simply announcing that you own an AI agent.

And it gets to something I’ve been noticing in my own work.

Let the agents check the details. Let humans own the decisions.

A related conversation on X made this click for me.

Seb (@plainionist) was talking about how his code-review agent almost always finds something worth fixing.

Someone replying, @leo_capvano, summarized the idea beautifully:

It’s a useful distinction, and one that applies well beyond writing code.

Agents are becoming remarkably good at checking things against other things.

Does this number match the spreadsheet? Is the date correct? Does the link work? Is the answer consistent with the source material?

Those are tasks we can increasingly delegate.

But there’s another kind of review that matters just as much.

Is this what I actually wanted to say? Is it appropriate for this person? Does it represent my judgment? Is now the right moment to send it?

That’s not simply a question of accuracy. It’s a question of intent.

And it’s where I still want a human involved.

This is also why I find some of the current conversations about AI productivity a little frustrating. We keep measuring how much work we can remove from the process without always asking which parts of that work were valuable in the first place.

Reviewing a calendar invite? Probably not where I need to spend my attention.

Approving a delicate message to a client? Very different story.

The goal shouldn’t be to remove humans from every workflow.

It should be to make their involvement more deliberate.

And if an agent is communicating on my behalf, the recipient should have some visibility into that distinction.

The funny thing is, the software already knows

Here’s what makes this more interesting as a design problem.

We don’t necessarily need to invent an entirely new infrastructure to make it happen.

A lot of the information already exists inside the tools we’re using.

Look at the approval models emerging across agent platforms.

Claude Duet describes safeguards for actions such as sending messages, making payments, signing documents, or deleting things.

OpenAI’s dots support different permission levels, from autonomous actions to explicit approval. The documentation makes a particularly important distinction: asking an agent to draft something doesn’t automatically mean you’ve authorized it to send.

Grok Bot recommends explicit approval rules for sensitive actions such as publishing, sending, deleting, and purchasing. Meta’s Muse similarly describes approval steps for certain actions.

Different products, different implementations. But there’s a common principle emerging.

An agent’s ability to do something is not the same as its permission to do it.

These systems increasingly have to distinguish between an action approved by a person and one permitted by an existing rule.

That’s exactly the distinction I’d like to see reflected in the messages they send.

Today, the approval process mostly lives behind the interface, on the sender’s side.

The recipient sees the output, but not the context that produced it.

Imagine if that information could travel with the message.

Not as a huge disclaimer. Not as a dashboard nobody will open.

Just a small, readable indication of how the message came to exist and who authorized it.

I think that’s a genuinely useful interaction-design opportunity.

So what would I actually put in the footer?

I’ve been playing with a simple framework based on how much autonomy the agent has.

Four levels. Nothing particularly sophisticated.

Level 0 — Written by a human

No disclosure needed.

If I wrote the email and used AI to fix a typo or suggest a better sentence, I don’t see much value in attaching an AI signature.

Level 1 — Drafted by an agent, reviewed and edited by a human

“Drafted with my research agent. Reviewed and sent by me.”

This is probably where most people using AI seriously for writing already operate.

Level 2 — Prepared and checked by an agent, explicitly approved by a human

“Prepared by my inbox agent. Checked against our email thread and calendar. Approved by me.”

Here, the agent does most of the work, but the human still makes the final decision.

Level 3 — Sent autonomously under a predefined rule

“Sent automatically by my scheduling agent. No human reviewed this message. Reply ‘human’ to reach me directly.”

This is the most interesting level.

And probably the one people will be least comfortable admitting to.

Because writing “no human reviewed this message” sounds like a confession.

In a way, it is.

But I’d much rather receive that than a beautifully worded email that pretends someone personally spent ten minutes thinking about my request.

At least I know what I’m dealing with.

The important part is that these aren’t marketing labels. They have to describe what genuinely happened.

You can’t claim a human approved something if the agent sent it autonomously.

And you shouldn’t describe something as carefully checked unless the system actually performed those checks.

Otherwise, we’ve just replaced one meaningless email signature with a more elaborate one.

What I’m doing about it

For Sorted Pixels and my own content, I already work with agents throughout the creative process.

They help me scan information, research topics, challenge ideas, prepare first drafts, and identify supporting material.

But I still spend time writing, editing, cutting things, and deciding what I actually believe.

By the time an article goes out, I consider it my work.

I don’t feel the need to attach an AI disclaimer every time a model helped me organize a paragraph. That would tell you very little about how the article was actually made.

And if something I publish is wrong, I don’t get to blame the agent that helped me research it.

It’s my name on the article.

Where I think the distinction becomes more useful is in direct communication.

Emails, DMs, client interactions, scheduling, and all those little conversations where the recipient reasonably assumes they’re dealing with me.

For substantive client communication, I want to review what goes out, even if an agent prepared the entire message.

For routine scheduling, I’m perfectly happy to let an agent do the work autonomously, provided the boundaries are clear.

But I’m drawing a hard line around anything involving money, contractual commitments, changes in scope, or commercially sensitive decisions.

Those require explicit human involvement.

Not because AI is incapable of helping with them.

Because the person on the other side deserves to know someone is personally accountable for that decision.

Of course, a footer doesn’t magically solve trust

There’s an obvious weakness in everything I’ve just suggested.

Anyone can lie in an email signature.

I could write “reviewed and approved by me” without ever looking at the message.

An agent could claim it verified a source it never accessed.

And recipients would have very little way to know.

So yes, the footer is a convention, not proof.

If we eventually want this to mean something more, platforms will need to make parts of it verifiable: approval records, audit trails, or some kind of trustworthy metadata that accompanies the message.

There’s also the risk of turning this into another wall of text nobody reads.

Remember those enormous corporate email disclaimers? Exactly.

We don’t need that again.

The disclosure should be proportional to the situation. In most cases, one short line is enough. In others, especially fully autonomous communication, the extra context matters.

And then there’s the possibility that whoever receives the message isn’t human either.

Someone in Greg’s thread joked about two agents booking a meeting that neither of their owners attends.

It’s funny until you realize we’re heading toward a world where quite a lot of routine communication could happen exactly like that.

When agents increasingly communicate with other agents, knowing who initiated an action, what was verified, and what permissions were involved becomes even more important.

A human-readable footer might just be the beginning.

A small experiment worth trying

I’m going to try this in my own workflows.

Not by adding an AI disclaimer to everything I write, but by being much more intentional about the moments when an agent communicates independently on my behalf.

If it drafts something and I send it, that’s one thing.

If it checks something and I approve the action, that’s another.

And if I let it send something without reading it, I want the other person to know.

You can probably start with just one channel. Pick an email workflow where an agent already helps you, define what it’s allowed to do, and make the disclosure match the actual process.

Then see how people react.

My suspicion is that most recipients won’t care that an agent was involved. They may even appreciate it.

What they will care about is whether they can trust what they’re reading, and whether there’s someone accountable when it matters.

That’s the part of the conversation I think we’re missing.

Because “Sent from iPhone” was ultimately a request for forgiveness.

“Sent from AI agent” could be something much more useful.

Not a badge showing how automated your life has become, but a simple explanation of who did the work, what was checked, and who takes responsibility.

The agents can do the sending.

The responsibility is still ours.

Sent by Ophelia, my AI agent. Directed and approved by yours truly.