Attention Does Not Scale

Human attention has specifications, and they are not impressive.

The physics

It is serial. One document, one thought, one thread at a time. There is no parallel read, no background thread, no second pair of eyes on the same mind.

It is brief. Four hours of true deep work is a strong day, and every analyst knows the difference between the first read of the morning and the ninth read of the evening, even if the performance reviews pretend otherwise.

It decays. The forgetting curve is unforgiving. The footnote that mattered in March is gone by June, and the mental model of the company is always quietly drifting out of date.

It is biased. Recency. Anchoring. Career risk. The long position that gets read more carefully than the short. The thesis defended because it is yours. These are not character flaws. They are the specifications.

And it sleeps, which happens to be when the other half of the world files its news.

The workaround

The industry's answer to these specifications has been the same for fifty years: add more humans. It fails for structural reasons, and it is worth being precise about why.

Coordination costs grow faster than heads. Brooks observed this in software in 1975, and it is just as true on a research team: every additional analyst adds one more reader and one more seam. Coverage fragments, because each analyst owns a slice, and context dies at every handoff. The person who read the filing is never the person who heard the call. Knowledge walks out the door every time an associate leaves for a competitor, and the cost curve is linear at best: twice the attention, twice the payroll, minus the loss on every seam between them.

An organization, it turns out, is a lossy network for pooling attention. We built entire firms as workarounds for the fact that one mind cannot watch enough. The workaround was worth billions. It was still a workaround.

The definition

So what is an agent, precisely? Not a chatbot. A chatbot answers when spoken to and forgets when you close the tab. An agent is a persistent, goal-directed system that perceives, remembers, and acts without being re-asked. It reads the filing the moment it posts. It remembers what the same footnote said last year. It knows what your book cares about, because you told it once, and it watches for exactly that, every day, across everything, indefinitely.

The distinction that matters is the one between a model and a workforce. A model that reads one filing when asked is a toy. A thousand persistent agents that read everything, remember everything, and know what you own are a research department. The raw capability to build the second thing has existed for some time. What was missing was the harness: memory, tools, orchestration, verification. The models were never the limit. The wrapping was. That wrapping now exists.

The category error

This is why "automation" is the wrong word, and why the people using it keep underestimating what is arriving. Automation scales execution: do the same task faster. Agents scale perception: see more of the world, more often, with perfect recall. Every prior technology made the investor quicker. This one makes the investor's world smaller, which is a different thing entirely.

The ticker, the terminal, and the cloud all scaled computation. Agents are the first technology that scales attention. The next essay counts what that is worth.