Lee Robinson shares his experience migrating the cursor.com site from a headless CMS to an architecture based on plain code and Markdown files. This transformation, completed in just three days, cost 260 dollars in AI tokens across 344 agent requests.
The main objective was to eliminate the accumulated complexity that the CMS had introduced into the project. Robinson identifies five major sources of hidden complexity in modern CMS architectures.
User management constitutes the first point of friction, with multiple account systems requiring separate access management. Change previewing represents the second problem, involving complex draft-mode systems with additional authentication layers.
Internationalization forms the third source of complexity, with laborious processes to manage multilingual variations. CDN delivery and assets constitute the fourth issue, with costly usage-based pricing: the company had spent more than 56,000 dollars on CDN since September.
Finally, dependency bloat represents the fifth problem, where over-abstraction makes the code difficult to maintain and modify.
The results of the migration are significant. Beyond the modest cost of 260 dollars, the project generated 67 commits representing 43,000 lines added against 322,000 removed. Build times were cut in half. CDN cost savings amount to thousands of dollars.
Robinson formulates a central observation: "the cost of abstractions with AI is very high". Coding agents work better with simple, direct code than with complex abstraction layers. Investing tokens to eliminate unnecessary complexity proves not only profitable but self-funded by the operational savings achieved.
This experience illustrates a paradigm shift in the evaluation of software architectures. With coding agents capable of performing massive migrations quickly, the cost-benefit calculation of technical debt changes. Removing abstractions becomes economically viable when the alternative is maintaining complexity that slows down both humans and AIs.
The case also demonstrates that agents perform better on streamlined codebases, creating a virtuous circle between simplification and AI-assisted productivity.