In 1934 and 1935, Herbert Simon, then an undergraduate at the University of Chicago, did a field study in his hometown of Milwaukee. His subject was how the city ran its public recreation facilities, which two agencies managed together: the school board and the city public works department.
The heads of the two agencies seemed to agree on what the program was for. Neither seemed to be building an empire. Yet they argued, over and over, about one thing: how much money should go to physical maintenance, and how much to play supervision.
Simon’s economics books had a clean answer for that. Spend the next dollar wherever it does the most good, and keep shifting until the last dollar on each side does equally well. Decades later, in his Nobel lecture, he put the question this way: “Why did they not, as my economics books suggested, simply balance off the marginal return of the one activity against that of the other?”
The answer took the textbook apart. They didn’t, Simon concluded, because “intellectually, they couldn’t.”
Two playgrounds in one park
There was no way to measure what a dollar of maintenance or a dollar of supervision produced. Each manager had a rough picture instead, and the pictures didn’t match. To the public works administrator, Simon wrote, a playground was “a green oasis in the crowded gray city.” To the recreation administrator, it was “a social facility, where children could play together with adult help and guidance.”
Organization theorists, Simon noted, call this subgoal identification. When a big goal can’t be tied to actions, decisions get judged against smaller goals that can.
Nobody can weigh every option, so we decide with the pieces we can see and measure.
The textbook decision maker, as Simon described it, needs to know all the alternatives, the consequences of each, and how to compare consequences that look nothing alike. Real people have to search for alternatives, can compute only so much, and face an uncertain world. A shopper in a supermarket aisle with forty cereals doesn’t compare them all on price per gram, sugar and fiber. The shopper reaches for the known brand, or the one on sale.
Simon listed the moves people use to make impossible decisions possible. Look for a choice that is good enough rather than the best one, which he called satisficing. Swap a big abstract goal for smaller ones you can observe and measure, which is what the two Milwaukee officials had done. And split the decision among specialists, coordinated through lines of communication and authority. All of these, he said, fit under the heading “bounded rationality.” Organizations themselves, in his view, were “machinery for coping with the limits of man’s abilities to comprehend and compute.” The idea ran through his 1947 book Administrative Behavior, and in 1978 he won the Nobel Prize in economics “for his pioneering research into the decision-making process within economic organizations.” His Nobel lecture began its account of the theory in Milwaukee.
The same narrowing showed up later in executives. In a 1958 study, Simon and a colleague named Dearborn used data on how businessmen read one case from a business policy casebook and what they saw as the company’s main problem. Sales and accounting executives mostly saw a sales problem. Manufacturing executives mostly saw a problem of internal organization.
Too many résumés to read
Imagine you’re hiring a product designer. You have 180 applications and one afternoon before the shortlist is due. What you want is someone who will do excellent work on your team. That can’t be read off a résumé.
So you use what you can read. Five years of experience. A portfolio link. A company name you recognize. By six o’clock you have twelve names, and they look a lot alike.
Somewhere in the other 168 is the designer with four years at a company you’ve never heard of, whose portfolio you didn’t open. The rule that cut the pile to twelve also decided, without anyone looking, that you would never find out.
Agents with a budget
AI agents run into the same bounds in a different form. Everything a model can consider at once has to fit in its context window. In July 2025, researchers at Chroma tested 18 language models, including GPT-4.1, Claude 4 and Gemini 2.5, on deliberately simple tasks. Performance dropped as the input got longer, often in uneven and surprising ways.
Anthropic’s engineers, writing in September 2025, borrowed the vocabulary of Simon’s economics books. Context, they wrote, “must be treated as a finite resource with diminishing marginal returns.” Their fixes echo his third move. A lead agent hands focused jobs to sub-agents, each of which may explore with tens of thousands of tokens and then send back a summary of often 1,000 to 2,000. When a conversation nears the limit, it can be compacted: summarized, and restarted from the summary.
The risk sits in the summary. Overly aggressive compaction, Anthropic warned, can lose “subtle but critical context whose importance only becomes apparent later.”
You can watch for that. Early in a long session with an agent, give it three constraints. Near the end, ask it to list them. If one is missing, the agent isn’t reliably retrieving it. That doesn’t prove the constraint has left its context, and a complete list doesn’t prove the agent will follow all three. Either way, open the session or its summary and see whether the constraint is still there. Then restate the conditions that matter at the top of your next message, instead of leaving them thirty turns back.