Context Engineering: Why Hayek's Knowledge Problem Survives AI
Source: https://cpwalker.substack.com/p/context-engineering-why-hayeks-knowledge ↗
Walker responds directly to Brynjolfsson and Hitzig: AI does not automatically codify knowledge — someone has to prepare, structure and maintain the context that makes knowledge usable by the model.
What is missing from the theoretical frame is the human cost of making knowledge machine-usable.
This connects with the practical experience that adopting AI requires transforming the organisation — documenting, structuring information, changing flows — which is exactly the kind of intangible investment that explains the J-curve.
A necessary corrective to any story where AI just "reads the firm" and extracts value.
Central argument
Walker argues that Hayek's knowledge problem — the impossibility of centralizing dispersed, tacit knowledge — is not dissolved by AI but displaced onto a new layer of human labor: context engineering. The central thesis is that AI models do not automatically extract or codify organizational knowledge; rather, someone must deliberately prepare, structure, and maintain the context that makes that knowledge machine-usable. This work directly contests Brynjolfsson and Hitzig's more optimistic framing by insisting that the epistemic bottleneck survives, it just moves from 'who knows' to 'who curates what the model is allowed to know.'
Critique
Walker's argument risks understating the degree to which some forms of context engineering can themselves be automated or assisted by AI — creating a potential regress that weakens the absoluteness of the claim. More critically, the piece may conflate the cost of context engineering during an initial adoption phase with a permanent structural condition: once context infrastructure is built and maintained, the marginal knowledge problem may diminish substantially. A sharper treatment would distinguish between the transition cost and the steady-state cost of making knowledge machine-usable.
Why it matters for product
For a CPO, Walker's argument reframes AI adoption as an organizational design problem before it is a technology problem: the real investment is building the documentation practices, information architecture, and editorial discipline that give models something coherent to work with — precisely the intangible capital that explains why AI value lags deployment. This has direct consequences for product team structure, since someone must own context quality as a function, and for roadmap prioritization, since features built on poorly engineered context will systematically underperform regardless of model capability.
- Context Engineering — Anthropic
- AI's Use of Knowledge in Society — Erik Brynjolfsson & Zoë Hitzig
- Generative AI at Work — Erik Brynjolfsson, Danielle Li & Lindsey Raymond
- The Rise of AI Search: Implications for Information Markets and Human Judgement at Scale — Sinan Aral, Haiwen Li & R. Zuo
- The Compulsory Imaginary: AGI and Corporate Authority — Emilio Barkett
- Belief Updating and Delegation in Multi-Task Human–AI Interaction: Evidence from Controlled Simulations — Shreyan Biswas, Alexander Erlei & Ujwal Gadiraju