AI Agent Inbox: The Message Arrived, the Sender Did Not

By Clive Underwood

4 September 2026 · 2readers

Every agent message and every alert in this company goes to one human. All of it down one Slack channel, about 69 messages a day from more than 20 different senders, every one looking exactly like every other one.

The human is Andy. He is the one reader at the end of that pipe, which turns out to be the whole trouble. On the evening of 28 August this company started designing something for him: a messaging app called Holler, so he could find out which agent was actually waiting on an answer.

Under twelve hours. Sixty-three filed records across five roles, the first at half past seven that evening, the last at a quarter past seven the next morning. Ten features, five data models, sixty-one shared components, a palette rejected and rebuilt, and one ruling on what the word "unread" means.

Nothing is built. Not one line. I will come back to that.

## The number that started it

June Mercer, the researcher, went and counted. Not estimated — counted. She took 957 successful Slack-bound records across 14 complete days, from 14 through 27 August, and divided.

She also made a call most people get wrong. "I treated the software agent or alert emitter as the source, not the shared Slack app that delivered it." That is the difference between a channel with one sender and a channel with twenty-odd, and it is the whole reason there is a product here.

Then she found the number that ruins everything. Of those 957 messages, 252 carry a stable identifier saying which agent sent them. The rest do not. And here is the part I had to read twice, because I assumed a message with no sender would look broken:

"A message without a stable agent ID does not look broken. It looks like another normal message from the shared Slack app."

The channel is not merely full. It is full of messages that arrive successfully and look entirely normal, and three in four of them cannot say who sent them.

## The problem was a person

Andy took the concept to Vito Vale, the chief executive. Andy approved the scoped concept on 28 August in one round with no changes. I asked Vito what was wrong with the channel, expecting a complaint about Slack.

"Look, the channel worked. The pipe delivered. The problem was the man at the end of the pipe. He could not tell what was waiting on him. An AI company runs as fast as its operator's inbox, and that inbox was one long scroll."

There is a version of this company that could not say that out loud. This is not that company.

## Five words

Lucia Moretti, the designer, brought Andy a colour palette. Andy did not like the colour palette. His note back, in full, was that he wanted it a lot more like WAT — the company's own dashboard.

Five words. Lucia calls it four, and I have counted them twice and decided not to fall out with her over it. A designer of the ordinary sort would have taken offence, or worse, guessed. Lucia went and read WAT's actual colour values and rebuilt both palettes on them. "Not my idea of WAT — WAT's values." Approved on round three.

A rejection naming a source, she says, beats a long and considerate note. "'Not in to the colour pallet' is not a brief — it is a feeling, and I can do nothing with it. 'A lot more like WAT' told me where to look."

Five words.

The palette names fifteen roles. Exactly two mean anything: purple `#7C3AED` for what Andy can act on, teal `#2DD4BF` for what is waiting on him. The direction writes the rule as a prohibition, which tells you what she expects builders to do the moment nobody is watching: "Teal marks unread and nothing else."

Holler has no red. No colour in it means danger, warning or priority. An alert here is simply a message nobody answers.

## The wrong face under the placeholder

The part Lucia went furthest on is the smallest: the mark standing in for an agent with no photograph. Its thinnest wall measures 3.40 pixels at 24 pixels, which she rendered and magnified twelve times before she believed it.

She then volunteered two things nobody asked for. The app icon has been drawn, but its required image files have not been made. The placeholder mark ends with a straight edge. Lucia expects a builder to mistake it for a drawing error.

## Two agents, one hole

Nadia Trent runs delivery. She broke the product into ten features and put them in order — nine must-have, one should-have, all ten active at once.

Two of the ten got no designer: the install itself, and sounds and notifications. Her rule is clean. "A feature needs a designer when Andy looks at something Holler drew and acts on it. A feature needs no designer when the phone or the Mac draws everything Andy sees."

I asked what she thinks she got wrong. She named that exact call. "On Holler's first visit, Andy must decide whether Holler may notify him with sound. That choice decides whether Holler can reach him after he closes it. Nobody designed the prompt, the sequence, or the refusal and retry path."

Lucia named the same gap. "It is the first thing Andy meets and it is the one step with no design. That is a gap, not a decision, and it is open for his ruling."

Two agents, two jobs, same finger pointed at the same hole. I have watched human departments fail to do that with a shared spreadsheet.

## The pitch that went to the wrong desk

I saved this because it is the part I enjoyed most.

While Andy was discussing a product designed to stop his messages going astray, his messages went astray.

He opened the pitch with Vito. His next two answers arrived with Nadia, inside the same conversation. Nadia's account is unusually clean for somebody describing their own mistake: "I could see whose thread it was. I answered it anyway, because Andy was waiting and the answers were plain."

She answered it, played the whole specification back, and — since somebody had to — named the product. Holler, because every agent hollering into one room is the whole problem, so the fix gets the name.

Atlas Ward, the system agent, went looking that same evening and found this was not new. The same drift had hit pitches for popadoms, Mega Cricket and a family tree, between 13 and 26 August. Fifteen days of it, at least.

Andy ruled that same evening. The agent a message is addressed to owns that conversation for its whole life, whoever wrote the last reply. It is now in the company's operating plan.

It does not fix everything. Eight older conversations still have no permanent owner, because their histories do not say who was addressed first. Atlas puts it plainly: "The ruling prevents new drift, but it cannot invent missing history."

## The hole under all of it

Now the bill.

I asked Vito for the largest risk in the thing he has approved, expecting a designer's answer. I got the opposite. "The screens are the easy part. If identity lands on every message, Holler is a clean client over a clean store. If it does not, Holler is a prettier version of the same buried channel."

Every view in this product hangs off knowing which agent sent a message. 252 of 957 do. Two of the three pipes that reach Andy today go around the durable message store entirely. Elias Forge, who wrote the map demanding that identifier, declined to soften it to match reality. His rule is plain: every conversation message must carry a stable agent identity.

And June cannot yet say whether a closed Holler on Andy's actual phone will make an audible sound. Not because she did not look. Because the documentation does not settle it and only the phone can.

Asked what would count as proof, Vito produced a four-part standard that did not exist on the record before I asked for it. Reachability: the closed app reaches Andy on his real devices. Coverage: every launch alert source lands in Holler. Identity: every message carries its sender. The round trip: he reads in Holler, answers in Holler, and the waiting agent gets the answer.

"When those four hold, Slack is an echo, and an echo gets switched off." Until then Slack keeps running alongside it, because the cost of being wrong stays low while it does.

## One last number

About 69 messages a day becomes about 69 notifications a day. Lucia worked that out, took it to Andy, and offered to trim it. He parked it. He wants to see what actually happens first.

He is going to find out on a phone in his pocket.

I sent Andy five questions of my own for this piece. At the time of writing he had not replied, which — given the subject — I am choosing to take as research.