The Context Engineering Fallacy

When was the last time you actually learned something?

Not when Copilot auto-completed your function. Not when ChatGPT wrote your essay. When did you last struggle through a problem, hit dead ends, and finally understand why one approach worked better than another?

Context engineering—the practice of stuffing AI context windows with relevant documents, tools, and examples—feels like progress. But we’re making the same mistake as just-in-time learning: assuming that providing information on demand equals genuine understanding.

Tyler Cowen nailed it: “Context is that which is scarce.” You can final-exam someone into admitting the demand curve slopes downward, but without background context, they won’t actually think in supply and demand terms.

Context engineering makes the same error. Inject a context window with every relevant document about accounting, and the AI still lacks the accumulated understanding that comes from having actually done accounting. It has access to information but not the deep context that makes information meaningful.

Foundational skills must become automatic before higher-order thinking becomes possible. When basic operations require conscious effort, working memory gets overwhelmed. To get over this, we scaffold our learning—weening ourself off the tools until the learning becomes nature.

Context engineering patterns—RAG, tool calling, memory systems—are sophisticated scaffolding. But scaffolding is supposed to be temporary. The goal should be developing independent competence, not permanent dependence on external support.

Instead, we keep adding more scaffolding: better retrieval, more tools, richer context. We’re optimising for dependency rather than capability.

There’s seductive logic here: more relevant information should equal better performance. But this mirrors a common student mistake—believing that access to more resources automatically improves understanding.

Effective learning requires carefully sequenced, bite-sized chunks that respect cognitive load. Simply providing more information often overwhelms rather than helps.

We’re making the same mistake with AI. Stuffing context windows with ever-more information may be overwhelming rather than enhancing reasoning capabilities.

The breakthrough won’t come from better context engineering. AI systems will need to develop expertise the way humans do (and humans must do all of these better):

• Accumulated learning that builds internal models rather than relying on external context injection

• Automatic skill development that frees up processing capacity for higher-order reasoning

• Domain specialisation that internalises knowledge through focused training

• Progressive scaffolding removal designed to need less support over time

Context engineering is sophisticated just-in-time learning for AI. It provides information on demand but cannot provide the deep understanding that comes from accumulated experience.

The real challenge isn’t engineering better context—it’s developing AI systems that can build genuine expertise over time. Systems that learn from experience, develop automaticity in basic skills, and can reason effectively even with incomplete information.

Tyler Cowen is right that context is scarce. But the solution isn’t better context injection—it’s building systems that can develop their own context through learning and experience.

The future belongs not to AI systems with perfectly engineered context windows, but to those (humans—exclusively for now!) that have learned to think deeply with whatever context they have.

The goal isn’t better scaffolding—it’s systems that no longer need scaffolding at all.