The skills/MCP/harness stack is the consumption-side efficiency discipline of a metered utility (cf. Tymshare service bureaus, electrified factories). The constraints are all set by the supplier across the API price boundary:
- context window — technical
- rate limit — administrative
- the bill — economic
Design decisions track these prices, not abstract elegance — e.g. prompt-caching's differential pricing reshapes how you lay out context. "Architecture follows the tariff."
But pure extraction misses half the picture: the stack is bidirectional — it also rations the unpriced-but-bound resource, human verification bandwidth → two-scarcities-arbitrage. And it does structural hedging: a model-agnostic harness = multi-sourcing; local models = going off-grid. The durable core is the constraint-surface interface; userland tricks depreciate at model-release cadence.