Coding Agents and Complexity Budgets
· agents, product-design
Key Points
- Lee Robinson (Vercel VP) migrated Cursor.com from CMS back to raw code/Markdown
- Completed in 3 days, cost $260 in tokens (297M tokens, mostly cached)
- 344 agent requests, 67 commits, 43K lines added, 322K removed
The Problem with CMS + AI Agents
CMSs introduce hidden complexity:
- User management duplication
- Preview workflows
- Internationalization hurdles
- Asset delivery costs ($56K CDN bill)
- Abstraction bloat
The network boundary is costly — agents cannot grep or understand data trapped behind APIs.
The Core Insight
The cost of abstractions with AI is very high.
Every layer between AI and the thing it's working with = friction, cost, failure points.
| Abstracted | Direct |
|---|---|
| CMS API (auth, fetch, parse) | Read file, edit, save |
| 10+ tool calls | 1 tool call |
Why Abstractions Hurt AI More Than Humans
Humans can:
- Click through GUIs intuitively
- Remember documentation
- Handle unexpected changes
AI agents must:
- Make API calls for everything (tokens)
- Re-learn abstractions every session
- Parse responses, handle edge cases
The Tradeoff Flipped
| Era | Abstraction value |
|---|---|
| Pre-AI | Abstractions save human time (worth complexity) |
| AI era | Abstractions cost tokens + reduce effectiveness |
Implication
If an agent can work directly with the thing, don't put a layer in between.
With AI agents, reducing abstraction layers becomes economically rational. Complexity that seemed necessary now has a simple solution: "spend tokens."
My Takeaways
- This validates MindCapsule's design — just Markdown files, no database
- Architecture decisions need rethinking in AI era
- "Code as content" wins because agents can grep/read/edit directly
- The simpler the environment, the better agents perform
Related
- Domain-Specific Agents — same principle: reduce decision surface
- Cloud Infrastructure — where this insight lives in tech context