I’m writing Agency Is All You Need in public — every chapter ships as an essay, gets argued with by readers, then revised into the final book (Q1 2027).
Sonja’s laptop knows things about her that she would rather not know about herself.
It’s a Tuesday evening in January, a little after nine, kitchen table, tea going cold beside the trackpad. She has opened the folder she calls Weiterbildung — professional development — which is where good intentions go to be archived. The dashboards load one by one. “Data Analytics for Finance Professionals”: 11 percent complete. “Python Bootcamp — Zero to Hero in 42 Hours”: 2 hours and 14 minutes watched. A machine-learning certificate from a business school, €1,890, paid out of her own pocket because the corporate training budget “prioritizes role-relevant skills” and apparently her future is not role-relevant: 61 percent, abandoned in November when quarter-end closing swallowed three weekends in a row.
This is not who she is. Sonja — mid-thirties, insurance economics degree, ten years of doing everything right at a large insurer whose restructuring memos have recently started deleting the career ladder above her — finishes things. She finished her degree near the top of her class. She has a system for everything: her expenses, her half-marathon training, her mother’s medication schedule. And yet here is the evidence, in cheerful progress bars: eleven enrollments, one completion — and the completion was a mandatory compliance module.
She tries again, because trying again is what she does. Timer, twenty-five minutes, Python course, module three: a man with a soothing voice explains list comprehensions over lo-fi music. At minute nine she checks her phone — reflex, not decision — and LinkedIn shows a former colleague announcing his new role: Lead, AI Transformation. Rocket emoji. Forty-one likes.
She puts the phone face down and realizes something: she could let the soothing man talk for the remaining sixteen minutes, and at the end the progress bar would advance and she would have learned, as far as she can honestly tell, nothing. The course would count it. She would still not be able to write three lines of Python.
“Okay,” she says to the empty kitchen. “So that’s not learning. Then what is?”
It’s the first useful question she has asked in months.
The capability that renews the others
Every skill you have is a snapshot. It was accurate the day you acquired it and has been quietly aging since — and in a period when AI rewrites whole categories of work in years rather than generations, it ages fast. That gives one capability a special status: self-directed learning is the one that renews all the others. Not because it’s noble — because it’s the maintenance system. If it works, obsolete skills are a solvable problem. If it doesn’t, every skill you have is a countdown.
Note what this is not saying. It’s not the familiar sermon about lifelong learning, which in practice translates into exactly the folder Sonja opened: enrollments as absolution. The question is not whether you consume educational material. The question is whether you can, on your own initiative, take a capability you don’t have and reliably build it — without an institution structuring the process for you.
For most adults, the honest answer is no. And that is not a personal failing. It’s the predictable result of how we all learned to learn.
You never learned without scaffolding
Think back through your own education. Someone else decided what you would learn, and in what order — the curriculum. Someone else decided when — the schedule, the deadline. Someone else checked whether you had learned it — the exam. Someone else supplied the consequences that kept you going when motivation failed — grades, teachers, the cohort around you doing the same thing. That entire apparatus is scaffolding: external structure that substitutes for self-direction. The great achievement of twentieth-century education systems — I made this argument at length in Education is Broken — was to manage the agency problem so thoroughly that most people never had to develop the capacity to manage it themselves. The system carried you. You supplied compliance and effort; it supplied direction, pacing, feedback, and persistence.
Then you left. And the scaffolding did not come with you.
An adult who wants to build a new capability faces the same problems the school solved — what, in what order, at what pace, with what feedback, sustained by what — with none of the machinery. Most of us respond by buying scaffolding-shaped products: courses, apps, certificates. But a video course is not scaffolding. It supplies content and a sequence; it supplies no consequences, no feedback on your work, no one who notices when you stop. The progress bar measures exposure, not capability. The graveyard of abandoned online courses is not evidence that people are lazy. It is evidence that they were handed a tool that assumes a capacity the school system never asked them to build.
The clearest demonstration came from the first great wave of massive open online courses. The world’s best universities put their teaching online, mostly free — and completion rates settled in the single digits. More telling than the dropout rate was who finished: overwhelmingly people who already held degrees, already trained by years of successful schooling to structure their own effort. The tool that was supposed to democratize learning amplified an existing advantage instead. Burn that pattern into memory, because it now repeats with AI tutors, which are more capable than any MOOC and just as indifferent to whether you show up: tools that require direction, persistence, and judgment advantage the people who already have direction, persistence, and judgment. The technology is neutral. The scaffolding is gone. What remains decisive is you.
What learning actually is
If exposure isn’t learning, what is? The research tradition here is unusually solid and unusually ignored. Barry Zimmerman and colleagues spent decades studying self-regulated learning — how people who successfully teach themselves actually do it — and found a cyclical structure: forethought (a specific goal and a plan), performance (doing the work while monitoring yourself), and self-reflection (judging the result, adjusting the approach). Albert Bandura’s work on self-efficacy supplies the engine: your belief that you can learn something is built mainly from mastery experiences — concrete evidence of your own progress. The belief and the capability grow from the same loop, or wither together.
Three consequences follow, and each contradicts how most adults try to learn.
Learning is production, not consumption. Watching someone code is not coding; reading about negotiation is not negotiating. Worse, consumption produces a feeling of fluency — the material makes sense as it flows past — that is easily mistaken for capability. The feeling is real; the capability is not. The test is always the same: can you produce something you couldn’t produce before?
Learning requires feedback, and feedback requires exposure. You need something that pushes back — a compiler, a client, a market, a colleague who reads your draft. Without an external check, self-assessment drifts toward comfort. This is the part adults avoid most, because feedback at forty stings in a way it didn’t at fourteen: you have a professional identity now, and being visibly bad at something feels like a breach of it. It isn’t. Being bad at things is the entry fee. There is no path to new capability that routes around the clumsy phase.
Learning is built through cycles, not events. Not one heroic bootcamp but repeated turns of challenge, feedback, and reflection, each slightly beyond current ability. And every loop you run does double duty: it builds the skill, and it builds the self-direction that makes the next skill cheaper to acquire.
The loop: build your own scaffolding
Since no one will structure your learning for you, build a small structure yourself. One loop, five parts.
1. Goal. Not a subject — a capability, stated so a stranger could verify it. “Learn data analytics” is a mood. “In twelve weeks, a script I wrote answers a question my team currently pays someone else to answer” is a goal. Size it in weeks, not years: self-belief is built from evidence of progress, so early goals should be sized to generate evidence.
2. Project. One concrete thing you will make, ideally connected to your real life or work, because reality supplies both relevance and feedback. The project doesn’t decorate the course — the project replaces the course as the organizing structure. Content (videos, books, an AI tutor) gets pulled in when the project demands it, in the order it demands. This inverts the consumption model: you apply first and learn what the application forces you to learn. Slower per fact, enormously faster per capability.
3. Feedback source. Decide in advance what will push back. Domains where the work itself gives feedback (code runs or it doesn’t) are a gift; elsewhere, arrange it — a person who reviews your output, a small audience, a real deadline. An AI tutor belongs here too, used correctly: not as a lecturer but as a sparring partner — critique this, quiz me, make me explain why, don’t accept my first answer. Used that way, it’s the most patient feedback source ever built. Used as an answer machine, it’s the fluency illusion with better production values.
4. Reflection. Fifteen minutes, weekly, written. What did I actually produce? Where did I get stuck — knowledge gap or avoidance pattern? What changes next week? This is the phase everyone skips, and skipping it turns the loop into a treadmill: motion without steering.
5. Next goal. The loop’s output is its own next input. When the project ends, the reflection tells you what’s next — deeper, adjacent, or a deliberate stop because the capability is good enough. That, too, is a legitimate outcome; navigation includes deciding where not to go.
Goal → project → feedback → reflection → next goal. It fits on an index card. What it replaces is an industry.
The honest hard parts
Everything above is straightforward. What makes it hard is not conceptual.
Energy after work is real, and willpower is the wrong tool. After eight hours of meetings, the version of you that enrolled is not the version on the couch. Design for the tired version: sessions on protected mornings if you can; ruthless reduction of startup friction (project open, next step written down from last time); sessions small enough that starting is easy. The self-determination research of Deci and Ryan points to the deeper fix: durable motivation feeds on autonomy, competence, and relatedness — a project you chose, sized to show progress, with at least one human who knows you’re doing it.
Your employer will not do this for you. Corporate training optimizes for the company’s current needs, compliance above all. Learning aimed at your next capability — especially one pointing beyond your current employer — is yours to fund and schedule. Unfair, and the situation. Treat the loop as personal infrastructure, like your savings.
And one caveat: the loop assumes a baseline of bandwidth. Exhaustion, depression, a life in acute crisis — these consume exactly the internal resources self-direction runs on. If that’s your season, the answer is not a better learning system; shrink the loop to something trivially small, or pause it without shame. The ability to return matters more than any single sprint.
Sonja builds a loop
Three weeks after the Tuesday of the progress bars, Sonja does something that would have embarrassed the fast-track version of herself: she picks a project so small it feels almost beneath her.
Every month she builds the portfolio steering report — four hours of copying figures between systems. Her goal, on an actual index card taped inside a kitchen cabinet: In twelve weeks, a script I wrote produces this report, and I can explain every line of it. Under it, smaller: Tu/Thu 6:30–7:20, Sat 9:00–10:30.
She sets up an AI tutor with standing instructions she typed herself: You are teaching me Python for data work. Never write whole solutions. Give me the smallest next step, then make me do it. When my code works, ask me why it works. Quiz me weekly. The first Saturday is humbling — forty minutes on a file-path error. The tutor is patient in a way no colleague would be: “What do you think the error message is telling you? Read it out loud.”
“I have a master’s degree,” she mutters at the screen. Then she reads the error message out loud, and understands it.
Week four, her script produces one correct column total. It is the least impressive thing she has built in her professional life, and she photographs the terminal output for her friend Meike, who has agreed to hear a two-minute progress report every Sunday — one human feedback source, chosen because Meike asks blunt questions and doesn’t care about insurance. Week six, reflection notes: Stuck twice — once knowledge (date formats), once avoidance (tired = phone). Fix: laptop open on table before bed, next step on a sticky note. Week nine: Realized I’ve stopped watching courses entirely. Don’t miss them.
The report ships in week eleven — imperfect, hers. Four hours become twenty minutes. But the real product is the loop itself, warm and running: she knows her next goal before she writes it down, because the claims data has patterns in it the report never asks about, and she now knows enough to ask.
For the first time since the restructuring memo, the future feels less like weather and more like water she can move through. Not because she knows where she’s going. Because she has remembered how to steer.
Sonja is a constructed figure — a persona built to carry real mechanisms, not a case study of a real person.