Attention Is All You Need. The Machines Have Plenty.

In 2017, a team of researchers at Google published a paper that quietly rewired how the world builds artificial intelligence. Its title was a provocation: Attention Is All You Need. They were describing a mechanism by which a neural network learns to focus on what matters inside a vast sea of information. It became the architecture underneath almost everything we now call AI.

We are living through a profound shift in how organizations deliver software and systems. AI agents now write code, manage workflows, and execute multi-step processes at speeds no human team can match. As agents take over the execution, the humans responsible for these systems are being quietly repositioned from builders to overseers. In practice, that means the quality of every decision, guardrail, and course correction now depends on something we are investing very little in protecting: the depth of human attention to the systems we operate. And we are already eroding it from within.

The same organizations deploying agents are also the ones replacing deep analysis with summarized briefings, and those briefings with summaries of summaries, compressing understanding at every layer until what remains is the confidence without the comprehension. Beyond the enterprise, the problem compounds further. The institutions and technologies that shape how we think, learn, and process the world are pulling in the same direction. Attention is becoming scarce precisely when it needs to become more disciplined.

The machines have the attention. The question is what happens when we need ours back.

It is true: attention is all you need. In the future, we, the humans, will need it more than ever.

The Speed Trap

In March 2026, Amazon’s retail platform suffered a series of high-severity outages over the course of a single week, leaving millions of customers unable to check out, access their accounts, or see accurate pricing. When the internal investigation surfaced its findings, the culprit was an engineer who had followed advice from an AI agent. The advice was reasonable, confidently delivered and drawn from the company’s official internal documentation. The engineer had every reason to trust it, but documentation was out of date. The system failed at the precise point where human attention had quietly stopped flowing.

This is worth sitting with, because the failure is subtler than what most people imagine when they worry about AI. The agent reasoned correctly from a corrupted source. The corruption was a knowledge gap, accumulated gradually and invisibly in the space between how the system actually worked and what anyone had recently bothered to write down about it.

That gap has a name now. Researchers call it cognitive debt, the erosion of shared understanding across a system over time. The code works and the architecture holds, but the living knowledge of why it was built the way it was, which decisions were deliberate and which were expedient, where the boundaries are and why they exist, that knowledge lives in people. And people relax, the moment the system starts running itself.

There is a related and newer concept called intent debt, which describes the absence of externalized reasoning, the goals, constraints, and tradeoffs that both human engineers and AI agents need to work safely with a codebase. Technical debt makes systems harder to change. Cognitive debt makes systems harder to understand. Intent debt makes it difficult to know what the system is actually for. The three accumulate together, in the same direction, and they are all accelerating.

Technical debt makes systems harder to change. Cognitive debt makes systems harder to understand. Intent debt makes it difficult to know what the system is actually for. The three accumulate together, in the same direction, and they are all accelerating.

The underlying force driving them is straightforward. AI agents have made it possible to build software at a pace that fundamentally decouples construction from comprehension. A developer can describe what they want, watch it materialize, test whether it behaves correctly on the surface, and ship it. The system works, it passes the checks, and the team moves to the next thing. This is the rational response to a set of incentives that reward velocity above everything else, and the understanding comes later, if it comes at all.

Think about what it felt like to learn a city before GPS. You took wrong turns, noticed landmarks, and built a mental map through friction and failure until eventually the city became yours in a way that a printed map could never replicate. GPS removed that friction entirely, and with it the internal map stopped forming. You can still get everywhere, but you can no longer navigate without the device. Apply that to a codebase, a system architecture, an enterprise platform. The agent moves through it with confidence, but the team only forms a surface level understanding of the systems.

The Deeper Current

What happened at Amazon did not begin with AI agents. It began much earlier, and understanding where it began matters, because the solution people tend to reach for, more governance, better documentation practices, stricter review processes, addresses the symptom while leaving the cause untouched.

In 2010, Nicholas Carr published a book called The Shallows, in which he argued that the internet was physically reshaping the cognitive architecture of the people who used it. He was drawing on neuroscience, specifically on the brain’s neuroplasticity, its tendency to strengthen the pathways it uses most and allow the ones it neglects to atrophy. His argument was that the internet, with its structure of hyperlinks, interruptions, and rapid context switching, was systematically exercising the circuits built for scanning and skimming while the circuits built for sustained, linear thought were going quiet from disuse. The online environment, he wrote, promotes cursory reading, hurried thinking, and superficial learning, not because people are careless, but because that is precisely what the medium rewards.

A decade later, Maryanne Wolf, a cognitive neuroscientist at UCLA who spent her career studying the reading brain, sharpened that argument into something more urgent. Deep reading, Wolf argued, is not simply a faster version of skimming. It is a categorically different cognitive process, one that involves making inferences, drawing analogies, examining the truth value of an argument, and integrating new information with existing knowledge. When we skim, she wrote, we physiologically do not have time to think. The speed that feels like efficiency is actually the thing that forecloses comprehension. And the digital environment, by rewarding speed at every turn, was eroding the population’s capacity for the slower, harder, more generative kind of reading that deep understanding requires.

Both Carr and Wolf were describing a process that was well underway before short-form video existed. Then short-form video arrived, and the process accelerated past anything they had modeled. Research now consistently shows that heavy use of platforms built on fifteen-second content is associated with measurably reduced capacity for sustained attention, with the effect strongest among the youngest users, the people who will spend their careers operating and overseeing the systems we are building today. The mechanism is the same one Carr identified, only the dosage is higher and the feedback loop is tighter. The brain learns to expect a new stimulus every few seconds, and tasks that withhold that stimulus, reading a long document, tracing a complex system, sitting with an ambiguous problem, begin to feel not just difficult but genuinely uncomfortable.

What is important to understand is that each of these technologies, the hyperlinked web, the social feed, the short-form video, offloaded something from us. The web offloaded memory and navigation. The feed offloaded the work of finding what was worth reading. The short-form video offloaded the patience required to let an idea develop. Each felt like a gain, and in narrow terms each was. But the cumulative effect was a population progressively less practiced at the kind of sustained, focused attention that complex systems require from the people responsible for them.

AI agents represent a different category of offloading. Every technology before this one took over a task, writing, navigating, searching, scheduling, while the human retained the full picture and exercised judgment about what to do next. What agentic AI introduces is something more consequential: it takes over the building, the doing, and increasingly the recommending. It surfaces the next action to take, the next change to make, the next decision to consider. And when the system is both executing and advising, the human in the loop faces a subtler erosion than simple deskilling. They begin to make decisions they no longer fully understand, drawing on recommendations generated from a system whose inner workings they have stopped tracking. The judgment remains theirs on paper, but the comprehension underneath it has thinned.

This matters because humans remain responsible for the outcomes regardless of how much of the work an agent performed, and that accountability will grow more consequential as these systems become more capable. Sophisticated observability tools can surface anomalies and point toward where something broke. The judgment about whether the architecture was sound to begin with, whether the tradeoffs made six months ago still hold, whether the recommendation the agent just surfaced reflects a genuine understanding of the system’s intent, that judgment belongs to a person who has maintained enough depth of knowledge to exercise it well. The future will reward people who retain that depth.

…humans remain responsible for the outcomes regardless of how much of the work an agent performed, and that accountability will grow more consequential…

What makes this particularly striking is that enterprises are already moving in the opposite direction. Organizations deploying AI agents to accelerate delivery are simultaneously deploying AI summarization tools to manage the volume of information those agents generate. Decisions that once required reading a detailed architecture document now get made from a two-paragraph brief. The brief is generated from a longer report that fewer people read. The report was itself a synthesis of code, tickets, and system logs that nobody had time to examine in full. At each compression, something is lost, and what is lost tends to be precisely the kind of contextual nuance, the edge cases, the historical reasoning, the deliberate constraints, that surface as critical knowledge exactly when a system behaves unexpectedly. The very organizations that most need their people to understand their systems deeply are building cultures that make that understanding progressively harder to maintain.

The Compounding Crisis

The 2024 National Assessment of Educational Progress, the broadest measure of student achievement in the United States, recorded the worst reading scores in a generation, with the largest percentage of students ever falling below basic proficiency in both fourth and eighth grade. What the assessment board noted alongside the numbers was telling: these declines were accelerating at the exact moment when technological advancement was demanding more of students and future workers. Reading scores are a lagging indicator. They reflect what years of accumulated habits, media environments, and institutional choices have produced in a population. What the 2024 data shows is a generation entering the workforce with a measurably reduced capacity for the kind of sustained analytical thinking that the oversight of complex systems requires.

That workforce is arriving into organizations that are themselves compressing understanding at every layer. Enterprises are producing more information than ever and reading less of it carefully. Briefings replace documents. Summaries replace briefings. The people making consequential decisions are working from increasingly thin representations of increasingly complex systems. AI tools meant to help them move faster are in many cases the same tools generating those thin representations. Research on AI-assisted reading is already showing measurable reductions in comprehension when people engage with summaries rather than source material directly. The institutions meant to develop deep analytical capacity and the enterprises meant to apply it are both moving in the same direction at the same time.

This is the compounding crisis. The systems we are building are growing more complex and more autonomous. The organizations operating them are losing institutional depth. The people staffing those organizations have been shaped by a media environment that has spent two decades optimizing against sustained attention. And the tools those people use daily are now completing the circuit by removing the remaining friction from reading, thinking, and deciding. Every layer is moving in the same direction simultaneously.

What makes this moment distinct is the asymmetry between the cost of inattention and the speed at which that cost arrives. In traditional software development, a gap in understanding accumulated slowly and announced itself through gradual degradation, rising maintenance costs, mounting technical debt, a system that became harder to change over years. In an agentic environment, the gap can be invisible until an agent acts on it, at which point the consequence is immediate, confident, and already in production. The Amazon outage took less than a week to unfold. The knowledge gap that enabled it had been accumulating for months. That asymmetry will only widen as agents become more capable and operate across more consequential domains.

The people who will matter most in this environment are those who bring something that cannot be generated or summarized: a genuine, maintained, hard-won understanding of the systems they oversee. The ability to read an architecture document and notice what is missing. To trace a decision back to its original intent and ask whether that intent still holds. To sit with an agent’s recommendation long enough to evaluate whether the confidence is warranted. These are the capacities that close the gap between what a system does and what the people responsible for it actually understand.

Summary

There is an old observation about boats and shallow water. Speed is fine in the open sea, where the margin between the hull and the bottom is deep enough that small errors in navigation carry no consequence. The shallows are different. The same speed that felt effortless offshore becomes the thing that determines whether you arrive or run aground.

We are moving very fast, and the systems beneath us are getting shallower in the sense that matters most. The distance between a confident decision and a consequential mistake is compressing. It seems that the agents are capable, the tooling is sophisticated, and the velocity is real and somewhere in that picture, the human with their hand on the tiller needs to actually know these waters.

It is true: attention is all you need. In the future, we, the humans, will need it more than ever.

Endnotes

1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems, 30.
2. Wharton AI & Analytics Initiative. (2026, April 14). Governing AI Agents: What the Amazon Outage Reveals about Enterprise Risk. University of Pennsylvania. https://ai-analytics.wharton.upenn.edu/wharton-accountable-ai-lab/governing-ai-agents-what-the-amazon-outage-reveals-about-enterprise-risk/
3. Storey, M.A., et al. (2026). From Technical Debt to Cognitive and Intent Debt: Rethinking Software Health in the Age of AI. arXiv:2603.22106
4. Carr, N. (2010). The Shallows: What the Internet Is Doing to Our Brains. W.W. Norton & Company. (Revised 2020)
5. Wolf, M. (2018). Reader, Come Home: The Reading Brain in a Digital World. Harper. See also: Wolf, M. (2020). Building Deep Reading Skills in a Digital World. The Guardian. https://www.centerfordyslexia.ucla.edu/maryanne-wolf-building-deep-reading-skills-in-a-digital-world/
6. Haliti-Sylaj, T., & Sadiku, A. (2024). Impact of Short Reels on Attention Span and Academic Performance of Undergraduate Students. Eurasian Journal of Applied Linguistics, 10(3), 60–68. See also: Short-form Video Use and Sustained Attention: A Narrative Review (2019–2025). ResearchGate. https://www.researchgate.net/publication/397712802
7. National Assessment Governing Board. (2025, September 9). Nation’s Report Card Shows Declines in 8th-Grade Science and 12th-Grade Math and Reading. https://www.nagb.gov/news-and-events/news-releases/2025/declines-in-8th-grade-science-and-12th-grade-math-and-reading.html
8. Ju, Q. (2023). Impact of Generative AI on Learning Efficacy in Academic Reading Materials. Duke Co-Lab Grant Program. https://colab.duke.edu/blog-post/grant-update-impact-ai-tools-reader-comprehension/