I had the most productive day of my career six months ago, and I was exhausted by 2pm.
I had three AI agents running concurrently: one refining architecture documentation, one generating test cases, one drafting a product brief from a call transcript I had given it. I was reviewing outputs, redirecting, approving, catching mistakes, rewriting prompts, merging work. Nothing in my hands was producing any of this. I wasn’t writing a single line. And yet, by mid-afternoon, I could barely hold a complex thought.
The strange thing was what I felt tired from. Not typing. Not designing. Not the kind of deep-focus coding fatigue I had always known. Something else. Something closer to the exhaustion of managing a team of interns who are brilliant but unreliable, and who never, ever stop producing.
That’s when it clicked: AI isn’t making me faster. It’s making me think differently. And my brain had not been trained for this.
The Brain Has Always Rewired for New Tools
Before we get into what AI is doing to cognition specifically, it’s worth placing this in context, because the alarmist framing (“AI is destroying our ability to think!”) misses something important.
Humans have rewired for every major cognitive tool. When writing was invented, Socrates famously worried it would destroy memory. He was partly right: we did stop memorizing as much. But we also freed up mental resources to think at scales that pure oral tradition could never support. The printing press didn’t make us dumber. It shifted which cognitive muscles we built. The internet did the same: a 2022 Stanford study found that heavy internet users developed superior visual-spatial processing and faster information filtering, while researchers also observed reduced capacity for sustained, deep reading.
This is neuroplasticity doing exactly what it is supposed to do: reorganizing neural pathways to match the demands of the environment. The brain is not precious about which circuits it reinforces. It builds what you use, and it prunes what you don’t. Maryanne Wolf, a cognitive neuroscientist at UCLA, has spent years studying how the brain physically changes its reading circuitry based on whether you spend your time scanning headlines or reading novels. Use it or lose it is not a metaphor. It is neuroscience.
What’s different about AI is the speed of the shift and the nature of what’s changing. Every previous tool mostly amplified one specific output channel. A calculator makes arithmetic faster. A word processor makes writing faster. AI doesn’t just amplify one channel. It generates entire swaths of output, across domains, concurrently, faster than you can read.
The bottleneck flipped.
Multiple Intelligences, Remixed
In 1983, Howard Gardner published Frames of Mind, proposing that human intelligence is not a single unified capacity (the old IQ view) but a profile of at least eight distinct abilities: Linguistic, Logical-Mathematical, Spatial, Musical, Bodily-Kinesthetic, Interpersonal, Intrapersonal, and Naturalist. Harvard’s Project Zero, which Gardner co-founded, spent years demonstrating that these intelligences rarely operate in isolation. They form constellations that shift based on what you’re doing.
The reason this matters for AI: working with AI requires a different constellation than traditional knowledge work did.
Before AI, a developer’s primary output was code. The workflow was essentially Linguistic (translating intent into syntax) combined with Logical-Mathematical (designing systems, debugging logic). Writing drove everything. The bottleneck was your ability to produce through that narrow channel.
Now consider what working with agents actually demands:
Linguistic intelligence shifts from production to evaluation. You’re no longer the primary writer. You’re the editor and the director. The prompts you write are more like creative briefs than output. Getting a model to understand your intent requires a different kind of linguistic precision, less mechanical and more conceptual.
Intrapersonal intelligence becomes newly critical in a way it wasn’t before. Knowing your own cognitive state, recognizing when you’re fatiguing, when you’re rubber-stamping, when your judgment is getting sloppy, is now a professional skill. You are the quality control layer. If you don’t know yourself well enough to know when your System 2 is fading (more on that in a moment), the entire output pipeline degrades.
Spatial intelligence gets activated for something it never had to handle: managing multiple parallel workstreams. When you have three agents running, you hold a kind of cognitive map of where each one is, what they’ve produced, what needs review, what’s blocked. That’s a spatial and organizational demand your brain wasn’t exercised on before.
Gardner himself has noted that as AI systems handle more of the computational, pattern-matching tasks, other less traditionally valued intelligences may become more important. I think he’s right. But the transition is not automatic. The intelligences you need to develop for working well with AI are not the ones your education trained, or the ones your existing workflow reinforced.
You have to deliberately build them.
The Supervision Trap (This Is Where It Gets Dangerous)
Here’s the thing nobody tells you about working with AI at scale.
Daniel Kahneman’s most important insight in Thinking, Fast and Slow is the relationship between two cognitive systems. System 1 is fast, automatic, intuitive, pattern-matching. It’s the part of you that catches a typo while reading, that reads social situations instantly, that drives a familiar route on autopilot. System 2 is slow, deliberate, effortful, logical. It’s what you engage when you verify a calculation, when you parse a legal document, when you review code for security vulnerabilities.
System 2 is also metabolically expensive. It burns through glucose and attention at a rate that System 1 doesn’t. This is why sustained analytical work is tiring in a way that habitual tasks are not. Kahneman calls System 2 “the lazy controller” because it is constantly looking for an excuse to hand the work back to System 1.
Now add AI to this picture. An AI model generates work at something close to System 1 speed, meaning fast and confidently. It produces code, copy, analysis, plans, and decisions in seconds. Your job is to review that output. But reviewing it properly, catching the subtle logical error, the hallucinated API method, the plausible-sounding but wrong architectural decision, that requires System 2. Full engagement. No shortcuts.
The trap is this: AI produces output at a rate that exceeds the sustainable throughput of your System 2.
When System 2 fatigues, System 1 moves in. And System 1 is terrible at catching confident-sounding errors. Kahneman identified several cognitive biases that specifically make us bad at scrutinizing plausible-seeming information. WYSIATI (What You See Is All There Is) means System 1 builds a coherent story from the information in front of it without questioning what’s missing. The Anchoring Effect means you’re disproportionately influenced by the first answer you see, which in an AI workflow is always the AI’s answer. Availability Heuristic means you evaluate quality based on how easy it is to generate examples, and AI-produced work comes with its own built-in plausibility that greases the slide toward approval.
This is what’s happening when experienced developers take 19% longer on tasks when using AI tools, per the METR study published in July 2025. I explored what separates developers who get faster from those who slow down in my analysis of when AI coding tools actually help vs. hurt. Part of the slowdown is prompting overhead. But part of it is the vigilance cost: skilled engineers know they need System 2 engagement to verify AI output, and maintaining that level of scrutiny over an extended session is genuinely draining.
A 2025 study in Frontiers in Psychology found that while AI cognitive offloading can reduce certain stresses, it simultaneously creates new pressures through “erosion of introspection, over-reliance on algorithmic feedback, and anxiety from hyper-monitoring.” The automation paradox: mental demand and monitoring pressure increase even as physical task demands go down.
What does System 2 fatigue actually look like in an AI workflow? It looks like approving a function without reading it fully. It looks like accepting a “this looks right” feeling when reviewing a document. It looks like catching yourself scrolling through model output instead of reading it. The output quality doesn’t visibly collapse. It erodes quietly, one rubber-stamped decision at a time.
The Bandwidth Flip Nobody Is Talking About
In August 2019, Elon Musk made a claim that sounded outrageous at the time and now reads like a design specification. Human communication to a computer, he said, will sound to future AI “like very slow tonal wheezing, kind of like whale sounds.” What’s our bandwidth? A few hundred bits per second, maybe a few kilobits per second if you’re being generous. He’s been making variants of this argument for years. We are already cyborgs, he says, but with “pathetic interfaces.”
The observation, stripped of Neuralink salesmanship, is accurate. The human visual system processes something in the range of 10 million bits per second. You can read approximately 250 to 300 words per minute under normal conditions, faster with training. Your output, whether typing or speaking, runs at somewhere between 100 and 300 words per minute. Your eyes are a high-bandwidth input channel. Your hands and voice are painfully slow output channels.
Traditional knowledge work leaned hard on that output channel. Writing code, drafting documents, composing emails: the bottleneck was always production. You could understand things faster than you could express them.
AI inverts this completely.
When an AI agent generates a 2,000-word report, 80 lines of code, or a full product specification in under 30 seconds, the bottleneck is no longer production. It’s consumption. Your high-bandwidth input channel (reading) is now the critical path. The question is no longer “how fast can I write this?” It’s “how fast and accurately can I read, evaluate, and direct this?”
This sounds like an upgrade. In some ways it is. But it changes the nature of the cognitive work substantially. Reading for critical evaluation is not the same as reading for comprehension. When you’re reviewing AI output, you’re not just processing information: you’re simultaneously running a fact-checking process, a logic-verification process, a “does this match my intent” process, and a “what did the model probably get wrong that I need to check” process. That’s parallel cognitive processing on top of input processing, and it happens against a clock because more output is coming.
The MIT Media Lab published a study examining what they called “cognitive debt”: the accumulated cost of outsourcing cognitive tasks to AI. The finding that stuck with me was that the debt isn’t just about atrophied skills. It’s about the brain’s reduced experience with the specific kind of effortful engagement that produces durable understanding. You can review a thousand AI-generated paragraphs and walk away knowing less than if you’d written ten of them yourself.
Input at volume is not the same as understanding. This is the trap hidden inside the bandwidth flip.
AI Is Not a Tool. AI Is Work.
Here’s the framing I keep coming back to, because I think it explains the cognitive experience more accurately than anything else I’ve tried.
A hammer is a tool. You pick it up, you apply force, you put it down. The work is yours. The hammer doesn’t do anything you didn’t physically do with it. A spreadsheet is a tool. You input data, you write formulas, you get an output. Your logic, your structure, your error.
An AI agent is not this. When you give an AI agent a task, what you get back is work. Actual output that needs to be evaluated for correctness, completeness, and fit-for-purpose, exactly the way you would evaluate work produced by a junior employee. The agent has used tools (web search, code execution, APIs) to produce that work. But the work product is not yours in the way a hammer stroke is yours. It’s the agent’s output, which you need to supervise.
This distinction is not semantic. It has direct cognitive consequences.
When you supervise a team, even a small one, you operate in a fundamentally different cognitive mode than when you produce individually. You hold context across multiple parallel workstreams. You make judgment calls about prioritization and quality bars. You track what’s been delegated, what’s been returned, what needs review, what needs redirection. You develop a mental model of each team member’s strengths and failure modes, because that’s how you know where to look harder.
Working with AI agents at scale is this. Not because AI is a person, but because the cognitive demands of the role are structurally similar. And like managing a team of very fast, very confident junior people with specific blind spots, the quality of your output depends almost entirely on the quality of your oversight.
A 2025 paper on human-AI task performance found that explaining AI weaknesses to human supervisors improved performance outcomes significantly. What that says, translated from academic language: when you understand where your AI tends to fail, you become a better reviewer. The mental model of the agent’s failure modes is part of the skill.
This is a professional capability that has no direct analogue in the pre-AI toolkit. The closest thing is code review, where you’re reading and evaluating work rather than producing it. But even experienced code reviewers aren’t doing this across 15 concurrent workstreams while also directing further work from those same workstreams simultaneously.
Training Your Brain for the New Reality
The good news is that neuroplasticity works in your favor here. The brain really can be trained for this. But you have to train deliberately, because nothing about the default usage patterns of AI tools builds the right habits.
The first thing to build is System 2 hygiene. Recognize that your deliberate attention is a resource that depletes. Schedule AI review sessions rather than doing continuous ambient review. Set hard limits on how long you go without a break when doing intensive output evaluation. The research on decision fatigue is unambiguous: the quality of judgment degrades predictably over time when System 2 is engaged without rest. Work with that biology, not against it.
The second is to develop a mental model of your AI’s failure modes. Every model and every context has characteristic errors. Hallucinations follow patterns. Reasoning failures tend to cluster in specific types of tasks. The time you invest in understanding where your tool is likely to be wrong is time you get back in faster, more accurate reviews. You’re building a specialized form of intrapersonal intelligence: self-awareness about the system you’re directing, not just about yourself.
The third is structured review, not scrolling. There’s a meaningful difference between reading AI output as if you were reading a Reddit thread, scanning for things that catch your eye, and reading it the way a senior engineer reviews a pull request, with specific questions in mind and a systematic check process. The second is much harder but produces substantially better outcomes. The Frontiers paper on cognitive offloading found that passive cognitive offloading (letting AI do things and accepting results) correlated with worse outcomes than active cognitive engagement with AI output.
The fourth, and the one I find most counterintuitive, is to maintain some deep-production habits. Write some code yourself. Draft some documents without AI assistance. Not because AI assistance is bad, but because the production mode activates different cognitive circuits that reinforce understanding in ways that review alone does not. The MIT cognitive debt research points directly at this: the durable understanding that comes from generation doesn’t transfer from consumption, even careful consumption.
Concurrent agent management is a skill. It improves with deliberate practice. The people who are genuinely good at it aren’t the people who use the most agents. They’re the ones who’ve figured out how to hold clear intent across parallel workstreams, how to write precise enough directives that the agents produce reviewable output on the first attempt, and how to ruthlessly cut scope when the cognitive load of supervision approaches the cost of just doing the work themselves.
The People Who Will Thrive Are Not Who You Think
Neuroplasticity means adaptation is possible. Not guaranteed. Possible. The brain remaps toward what you actually practice, not what you intend to practice or what the marketing materials say you’re practicing.
If you use AI primarily in passive mode, accepting outputs without rigorous engagement, you will likely develop faster shallow-scanning habits and weaker deep-evaluation habits. Your throughput will look impressive on surface metrics. Your quality will quietly decay.
If you use AI in active mode, treating the output as work to be evaluated and directed rather than content to be consumed, you’ll build something genuinely new: a hybrid cognitive capacity that combines strategic direction with rapid critical evaluation, spanning more domains and more parallel workstreams than was possible before AI.
The people who thrive won’t necessarily be the fastest coders, the best writers, or the deepest subject-matter experts, though those things still matter. The people who thrive will be the ones who understood early that the real skill in the AI era is the management of AI work quality, that this is a cognitive skill requiring deliberate training, and who started building the neural infrastructure for it before everyone else.
Your brain is plastic. The question is what you’re shaping it toward.
Working with AI agents yourself? I’d love to hear your experience, especially if you’ve found supervision techniques that actually work. Find me on X or LinkedIn.
Keep reading: For the broader story of how AI got here so fast, see the two years that rewrote AI from 2024 to early 2026. And for the practical side, here’s when AI coding tools actually make you faster vs. slower.
Research for this article was conducted in February 2026. Sources consulted include Daniel Kahneman’s Thinking, Fast and Slow (Farrar, Straus and Giroux, 2011); Howard Gardner’s Frames of Mind (Basic Books, 1983), Harvard’s Project Zero on Multiple Intelligences, and his reflections on AI and recognition; the METR study on AI developer productivity (July 2025); MIT Media Lab’s research on cognitive debt and AI writing tools; Frontiers in Psychology on cognitive offloading and AI (2025); Elon Musk’s bandwidth argument from Business Insider (2019); and the Stanford / NetPsychology neuroplasticity research.