Coding Agents Turn CMS Abstractions Into Complexity Debt
Maps to joelclaw's agent-authored content flow: keep content where agents can grep, diff, preview, patch, and ship it through git.
Lee Robinson migrated cursor.com from a headless CMS back to raw code and Markdown in three days, using $260.32 in tokens, 344 agent requests, and 67 commits. The headline isn’t “AI made migration faster.” The useful part is sharper: agent workflows make old abstraction tradeoffs more expensive.
The clever bit is that Cursor didn’t just replace a CMS with vibes. It replaced separate user management, preview mode, i18n plumbing, CDN usage, bespoke content formats, and abstraction bloat with code, GitHub, pull requests, object storage, and agents that can inspect the whole thing. A CMS GUI used to be the friendly layer. In an agent-heavy system, that GUI can become the wall the agent keeps faceplanting into.
This is very relevant to joelclaw because the whole system already treats text, state, and workflow artifacts as things agents should be able to read and change. Obsidian, Markdown, Next.js, git, and durable workflow receipts are a much better substrate than opaque content blobs behind a SaaS UI. Not always, obviously. But when the operator is comfortable reviewing diffs, content as code stops being developer-hostile and starts being agent-native.
The spicy takeaway is the “complexity budget” framing. Storybook, CMS preview systems, draft mode, custom localization plugins, and CDN-backed asset managers can all be good tools. But if the thing mostly exists because humans needed buttons, and the new primary operator is a coding agent plus a human reviewing a PR, the stack deserves a fresh trial. Some abstractions are now just rent-seeking bullshit with a dashboard.
Key Ideas
- Lee Robinson frames the CMS as an abstraction whose cost rises when coding agents need to inspect, edit, and verify content.
- The migration moved cursor.com content back into raw code and Markdown, making changes flow through git and GitHub pull requests.
- Cursor used scripts to export CMS content, validate structure, convert it into repository files, and move media into object storage.
- The article calls out hidden CMS costs: duplicate user management, preview complexity, localization workflows, CDN and asset delivery fees, dependency bloat, and formats agents cannot easily grep or patch.
- The reported result was $260.32, 297.4M mostly cached tokens, 344 agent requests, 66 manual Tab changes, and 67 commits for a migration Lee initially estimated at one to two weeks.
- Static pages and PR previews beat CMS draft-mode complexity when the team can review diffs and share deployment previews.
- For joelclaw, this supports keeping publishable knowledge in agent-readable files and workflows instead of hiding it behind UI-only systems.