Veritas Silentii Conversations Across Time
Conversations Across Time
Conversation No. 4 · September 2026
Herbert Simon, the Nobel Prize, and a Warning About the Agentic Shadow
Personal AI agents may eventually create two worlds: the world they believe belongs to us, and the world they decide does not.
We will live in the first.
We may never know the second existed.
Nearly fifty years ago, Herbert Simon was awarded the Nobel Prize for pioneering research into how people make decisions. A few years earlier, he had given us one of the defining insights of the information age, in a line that most of us now take for granted:
"A wealth of information creates a poverty of attention."
As with many groundbreaking ideas, Simon's insight was elegantly simple.
Our attention is not unlimited.
As the volume and complexity of information and choices increase, attention becomes an increasingly scarce resource.
So, our attention must be managed.
And we manage it by simplifying…by constructing rules and heuristics that reduce an unmanageable world to a manageable set of choices.
That insight helped define the information world that followed. And an enormous economy grew around the scarcity Simon had identified.
- The internet made information nearly infinite.
- Google helped us find it.
- Amazon helped us navigate it.
- Social media learned to capture our attention.
- Advertisers learned to shape and buy it.
But there was a second half to Simon's argument that, in hindsight, was an unheard warning for us living a half century in the future.
Most everybody missed it.
In the same 1971 essay, Simon asked another elegantly simple question.
What is an information-processing system actually for?
Its purpose, he said, was not simply to bring a manager more information.
It was to reorganize the information environment so that less of your scarce attention had to be spent receiving it.
This was an answer that the information industry that followed would not have wanted to hear…spend less of our scarce attention time involved with systems like Google, Amazon, and social media. And they didn't hear it, until recently.
Information-processing systems, he wrote, would reduce the demands on our attention only if they absorbed and processed more information than they produced…if, in his wonderful phrase, they "listen and think more than they speak."
In the language of information theory, the system should do something fundamental: reducing the uncertainty and complexity of too much information into something we can act upon. But to accomplish that reduction requires something to be left behind.
Read that again with today's eyes.
Simon was not describing a search engine of the following half century.
He was describing something remarkably close to the AI agents that are emerging today.
Their value lies not simply in what they deliver, but in what they absorb on our behalf…and don't deliver.
They earn their keep largely by withholding.
For half a century, there was little reason for most of us to dwell on that distinction. The tools we built primarily helped us organize and navigate available information. They knew relatively little about who we were, what we wanted, or what we might become.
Now that is changing.
We are building systems that don't simply organize the world in front of us. They increasingly learn who we are in order to decide which parts of that world should, and should not, reach us at all.
In that sense, we are finally building tools that do something remarkably close to what Simon prescribed…they listen and think more than they speak.
And the implications may prove even more consequential than we imagine.
Filtering Is the Job. Exclusion Is the Shadow.
If human attention is limited and information keeps expanding, an effective system cannot simply get better at finding, navigating and delivering more of it.
There is simply too much. It must filter. Condense. Prioritize. And this is the part we missed…more than select, it must exclude.
We are beginning to build intelligent, increasingly autonomous machines to select information on our behalf.
But to select on our behalf, they must also reject on our behalf.
Almost everyone today is focused on what agents will select.
Simon might have noticed the other side:
Selection itself is a curtain that casts a shadow of exclusion.
This is not necessarily a manipulation or bias. Instead, it is the inevitable consequence of the agent simply doing its job.
Simon had a name for the human limitation that makes all of this necessary. He called it bounded rationality.
We cannot examine every option, so we don't try. We search until we find something good enough, and then we stop.
He called that satisficing.
Our bounds…and our sense of what constitutes "good enough"…are what make our decisions possible at all.
But an agent that satisfices on our behalf is doing something fundamentally different.
It is applying its bounds and its sense of "good enough," informed increasingly by its model of who we are.
We are not simply delegating a task.
We are beginning to outsource our bounded rationality.
And the bounds are no longer entirely our own.
The Other Side of B2A
You can already see business waking up to one side of this.
Companies spent much of the last century learning B2C, then B2B.
Now a third audience is emerging: B2A…Business to Agent.
And in just the past few days, the idea has begun moving from just a future prediction to our current infrastructure.
Anthropic has introduced technology that allows retailers to build their own AI shopping agents. Shopify and Priceline are among the companies already working with this technology. The agents can search catalogs, compare products, personalize recommendations, assemble carts and connect customers to checkout.
Think about that through Simon's eyes.
The scarce resource is no longer simply the customer's attention. It's becoming the agent's attention.
A retailer may have thousands of products. A marketplace may contain millions. The customer cannot consider them all, which is precisely why the agent is useful.
But neither can the agent present them all to you.
It must decide which products deserve to survive the journey from abundance to attention.
Simon taught us that information abundance creates human attention scarcity.
Agentic commerce adds a different layer.
The agent does not face attention scarcity as we do. Its job is to manage abundance on our behalf by reducing a vast universe of possibilities to a small consideration set.
That creates a new kind of scarcity:
Admission to the agent's consideration set.
Thousands of products may be available. Hundreds may meet our needs. But only a handful will survive the agent's filtering and ever reach us.
In Simon's world half a century ago, businesses competed for scarce human attention.
In the agentic world, they may first have to compete for scarce access to the machine's consideration set.
That is already becoming measurable.
John Lewis recently reported that product searches coming through AI agents had risen from just 0.3 percent to 2.5 percent in a year…more than an eightfold increase. Retailers and measurement companies are beginning to talk about "agentic shelf visibility": whether a product gets considered by the machine at all.
The metaphor is revealing.
For generations, manufacturers competed for scarce physical shelf space.
Then they competed for scarce space on Google's first page.
Now they may increasingly compete for scarce space inside an agent's consideration set.
The shelf hasn't disappeared. It has become invisible inside the agent's software.
And unlike the physical shelf, the customer may never see what isn't on it.
That takes Simon's scarcity argument somewhere new.
When human attention was scarce, businesses competed to capture it.
When agents manage that scarcity for us, businesses must increasingly compete for machine attention before they can reach human attention.
But another development may be even more consequential.
Retailers are beginning to confront whether the shopping agent will belong to them or to someone standing between them and the customer.
That explains why preserving your direct customer relationship matters so much.
If another company's agent becomes the primary interface through which customers discover, compare and buy products, the retailer risks losing more than traffic or transaction data.
It risks losing influence over the consideration environment itself.
Simon called the facts, assumptions and constraints within which decisions are made decision premises.
In an agentic marketplace, whoever controls the agent may increasingly influence those premises: which products are examined, which attributes matter, which alternatives are compared…and which never appear.
The scarce resource is shifting from customer attention toward control of the filter that allocates customer attention.
And now the final step is beginning to emerge.
India is developing infrastructure that could allow AI agents to make certain purchases through its UPI payment network without requiring human approval for every individual transaction.
That moves the agent across another boundary.
- Search once required our attention.
- Comparison required our attention.
- Choice required our attention.
- Payment required our attention.
- Increasingly, for more of what we do, none of them will require our attention.
The progression is remarkably simple:
Help me find it.
Help me compare it.
Recommend one.
Buy it for me.
At each stage, the machine silently, and with our permission, absorbs another small piece of the attention Simon told us was scarce.
And at each stage, something else becomes scarcer:
Our direct encounter with the alternatives that were left out of the consideration set.
That is the hidden warning in Simon's insight.
The better the agent becomes at conserving our attention, the fewer possibilities we personally encounter.
Which means the great commercial question of the agentic age may not simply be:
How do we make sure the agent sees us?
There is another question on the other side:
What happens to everything the agent decides we don't need to see?
- For every product the agent shows us, what did it exclude?
- For every company it recommends, which companies disappeared before we knew they were alternatives?
- For every executive candidate who makes the shortlist, who didn't?
- For every investment opportunity it identifies, which possibilities never reach the investment committee?
- For every explanation it gives us, which competing explanations never enter the conversation?
The agent does not merely help us choose among possibilities.
It helps determine which possibilities become visible enough to be choices at all.
That is the beginning of what I think of as the Agentic Shadow World.
Two Worlds
Imagine that I ask my personal agent to find a hotel in Rome.
Today, I might search Google or a travel site and see twenty hotels. I know there are hundreds more. I can scroll farther. Change the search. Visit another site. Ask a friend.
I remain aware that the visible choices are only part of a larger world.
Now imagine that my agent knows me.
It knows where I have stayed before. What I like. What I spend. The neighborhoods I prefer. Whether I value a large room or a great restaurant. Perhaps even what my wife and I complained about after our last trip.
I tell it:
Find us a wonderful small hotel in Rome. You know what we like. Keep it under $500 a night. Book four nights.
Done.
Perhaps the agent examined 2,000 hotels, seriously considered twenty, and only three survived its final comparison.
I don't actually know. More importantly, I don't need to know. That is why I delegated the task.
But something fundamental has changed.
The hotels that disappeared did not lose because I rejected them. I never rejected them.
I never knew they existed.
The agent has created two worlds. The world it concluded was appropriate for me…and the much larger world it decided was not.
I live entirely in the first.
The second becomes my Agentic Shadow World.
Haven't We Always Had Filters?
That's a fair objection. After all, we have always lived inside other people's exclusions.
Editors decided what was news. Brokers decided which stocks to mention. Headhunters decided who made the slate. Librarians decided what sat on the shelf. Google decided what went on page two…and almost nobody looked.
So what is different now?
Three things. Visible. Plural. Edges.
The old filters were visible. You knew there was an editor, a broker, a headhunter. You could understand and question the filter.
They were plural. A different newspaper, a different broker, a different friend, multiple media sources, could show you a different slice of the world. The disagreements among them helped reveal the edges.
And those filters had edges you could step past. You could turn to page two. Walk to another shelf. Read another newspaper. Ask for a second opinion. The personal agent changes that.
Its filtering is effectively invisible by design, because its value lies precisely in the fact that we don't want to watch it work.
It can become personal, because it is ours…increasingly tuned to us and only us. There may be no competing filter whose disagreement automatically reveals the first one's blind spots.
And eventually it may become edgeless.
There is no page two. No defined beginning and end. Nothing farther to scroll.
The consideration set arrives. Invisible. Singular. And Edgeless.
And that combination forms a curtain that casts a shadow on other alternatives.
When the Exclusions Really Matter
Most of the time, this will feel like an extraordinary benefit. And most of the time, it is.
I don't need to examine 2,000 hotels, read 4,000 product reviews, or evaluate every possible route to the airport.
Simon understood precisely why filtering is necessary.
But the consequences change as the decisions become more important.
Suppose instead I ask:
Which three companies should we consider acquiring?
Or:
Who are the five strongest candidates to become our next division president?
Or:
What are the greatest competitive threats to this business?
Or, more personally:
Given everything you know about me, what should I do next with my career?
Now the exclusions matter. And that last question may eventually matter most of all.
Because our agents won't remain generic. They will study and learn our histories, preferences, ambitions, habits. And our risk tolerance.
What we buy. What we read. What we reject. What has worked for us before. And perhaps most consequentially, what people like us usually do.
That should make an agent enormously useful.
But it introduces a paradox.
The better the machine becomes at understanding who we have been, the more likely it may be to assume that who we have been is who we should continue to be.
Our past can become the evidence from which the future is constructed. And the experiences of thousands or millions of people judged to be "like us" can reinforce the boundaries.
The result may be recommendations that are extraordinarily well personalized…and remarkably conventional. Safe choices. Probable choices. Choices consistent with our history. Choices people like us usually make.
But lives and careers do not always advance through the most probable next step. Sometimes they advance because someone does something improbable…starts a business when the sensible choice is another corporate job or moves somewhere unexpected. Maybe changes industries or writes a book. Perhaps even takes a risk that fails…but through that failure meets someone, discovers something, or finds an opportunity that changes everything that follows.
The very experiences that transform us are often the ones our histories could not have predicted. And sometimes even the worst choices lead to the best outcomes.
That creates a danger deeper than bad recommendations.
An agent trained to understand us perfectly may become very good at recommending more of the person we have already been.
And if millions of us are increasingly guided only by what we've done before, or what people "like us" have historically done, personalization could produce an unexpected kind of sameness: individually tailored recommendations that quietly pull us toward the probable, the familiar, and the average.
Not because the machine tells us not to be ambitious or to take risks. But because the improbable possibility may never make the list.
Imagine a 55-year-old executive asking an agent what he should do next.
The machine knows his career, compensation, education, industry, geography and network. It recommends another executive position. Board work. Consulting. Perhaps teaching.
All perfectly reasonable.
But perhaps it never suggests buying a company. Starting one. Writing a book. Moving abroad. Entering an entirely different industry.
Not because those possibilities don't exist or that he couldn't succeed at them. But because the agent's understanding of someone like him caused those possibilities never to enter the agentic consideration set…they remained in the shadow.
Simon would have said that the agent satisficed on his behalf.
And its idea of "good enough for someone like this" was drawn largely from everything he had already said and done…and everything people like him had done before.
The machine didn't tell him: You can't do that. It did something much more subtle.
It never told him that he could.
And perhaps that is the most consequential shadow of all.
Great success usually requires overcoming the limits other people placed on us.
We may now have to learn to recognize the limits a machine infers from who we have been. At that point, personalization has acquired a shadow. The machine is no longer simply helping us navigate the world as it exists. It is quietly determining which portion of that world belongs inside ours.
Back to Simon
Simon argued that organizations influence decisions partly by establishing what he called decision premises: the facts, assumptions, values and constraints within which decisions are made.
Power does not always have to say: Choose X.
A subtler form of power establishes the environment in which X becomes the only sensible choice.
Now put an intelligent personal agent between us and an almost infinite world of information and alternatives.
It filters the environment. It determines what warrants attention. It withholds the rest.
And what survives increasingly becomes the world from which we make our decisions. And that is why the unseen matters.
Because the greatest consequence of exclusion may not be that we make the wrong choice.
It may be that some choices never become imaginable.
It may eventually lead us beyond Simon's economics of attention toward a much older question about legitimacy:
Who gets to establish the accepted boundaries of what matters, what is credible, what is possible…and ultimately, "the way things are supposed to be"?
But that is another Conversation.
The Simon Echo: Interrogate the Exclusions
For now, Simon leaves us with something practical.
Whenever you use an agent for an important decision, don't just accept its recommendation.
Ask what it left out.
For a CEO:
Which acquisition candidates did you exclude before giving me these three—and why?
For hiring:
Who didn't make this shortlist because of the criteria you used?
For strategy:
Which competitors did you dismiss that would become important if one of our assumptions were wrong?
For your career:
Which opportunities didn't you show me because of what you inferred about who I am?
There is a catch.
When you ask an agent what it excluded, you are asking the filter to audit itself. Its analysis of the shadow is drawn from behind the same curtain that cast it.
So, while the question is necessary, it isn't sufficient. Two habits can make it stronger.
First, ask a different agent the same question. The old world's protection was plurality. We can deliberately rebuild some of it. Where two agents disagree, the edges of each one's shadow may become visible.
Second, ask for the deliberate misfits:
Show me three options you would normally never recommend for someone like me—and tell me why.
Sometimes the answer will confirm the agent's judgment. Sometimes it may reveal something much more important:
Perhaps that "someone like me" was a lesser person than you actually are.
The point is not to distrust the agent. Quite the opposite. We will need these systems precisely because Simon was right. There is too much information and too little human attention to examine everything ourselves.
The danger is forgetting that the filter is making decisions too.
Nearly fifty years ago, Simon gave us the defining insight of the information age:
"A wealth of information creates a poverty of attention."
The agentic age may reveal its hidden inverse:
Automated attention creates a shadow.
When the decision matters, ask the machine to show you the world it left out.
Interrogate the exclusions.
What Inspired This Conversation
This Conversation draws on Herbert Simon's 1971 essay "Designing Organizations for an Information-Rich World," and on three developments from the past several days in agentic commerce: Anthropic's new blueprint for AI shopping agents, built with Shopify and Priceline among the launch partners; John Lewis's report that product searches arriving through AI agents rose from 0.3 percent to 2.5 percent in a year; and India's preparations to let AI agents complete UPI payments without transaction-by-transaction approval.