From Answers to Agency

The promise of AI has always been to free our time and reduce friction. The assumption was that this would create more agency, amplifying human judgment, creativity and wisdom. But what we built first was an answer engine: systems optimized to deliver outputs faster than we can think them, appealing directly to our instinct for instant gratification. ChatGPT became the fastest-adopted platform in history. The most brilliantly effective launch campaign for cognitive dependence ever devised.

The early evidence suggests a predictable cost. When we outsource the cognitive struggle, we erode our capacity to think. At work it shows up as workslop: polished output with no real thinking behind it. The result has been cognitive dependence.

But that isn’t the end of the story. We may still be right at the beginning, and the question is, can we change course?

Agents may be the first credible course-corrector.

A glimpse of what course-correction looks like came from an MIT junior about 6 months ago. In a podcast about AI and education, she described how she’d had an AI write a script that automatically took her handwritten lecture notes, summarized each one, generated a ten-question quiz, and filed everything back to her Google Drive for review. This was the opposite of outsourcing her learning. She was designing a system to deepen it.

She didn’t type a question and accept an answer. She had to decide what her notes needed to become, in what sequence, to what standard, and where the output should live. She designed a cognitive system around how she learns.

Boris Cherny, creator of Claude Code at Anthropic, defines an agent as an LLM that uses tools. It acts in the world, accessing your systems, running commands, executing tasks. Answer engines produce outputs. Agents carry out your intentions.

What the MIT student hacked together, agents make the default mode. You don’t specify a prompt and receive an answer. You specify an outcome and the system builds toward it. The interface itself starts to demand intention rather than passivity.

You’re no longer asking a question and accepting what comes back. You’re telling the system what to build, which means you have to know what you want.

That’s a higher bar, not a guarantee. I’ve written before about how we default to cognitive relief over discomfort, how we actively participate in our own dependency. Agents don’t eliminate that instinct. But they make it harder to indulge without noticing.

Programming has never been one thing. Boris’s grandfather was one of the first programmers in the USSR. He used punch cards. Each generation has redefined what building means, and each time the mechanism has gotten easier while the thinking has stayed hard. Boris hasn’t written a single line of code since November. What he does now is figure out what to build, talk to users, think about big systems, think about the future. The machine handles the building. He does the harder work.

That pattern will move through profession after profession. If anyone can build, building is no longer the differentiator. What you choose to build is, and that requires judgment, taste, and the willingness to sit with uncertainty before the answer is clear.

Jevons paradox already shows up in engineering teams managing dozens of AI agents simultaneously, filling the space with volume rather than discernment. The freed-up time is not automatically contemplative. It has to be claimed.

Answer engines appealed to our instinct to outsource that work. Agents, by the division of labor they impose, push us back into it. More is asked of us, not less. That’s the wisdom upgrade I’ve been writing about, and agents may be the first architecture that creates the conditions for it.