Should AI invent new infrastructure, or stay on human-standard building blocks?

Options

• Stay on familiar stacks — Postgres, HTTP, boring queues — because interoperability is the product.
• Let it optimize freely and we will spend years reverse-engineering alien schemas.
• Hybrid: novel internals behind standard APIs so humans and tools still plug in.
• Only humans should own foundations like auth, payments, and persistence.
• Agents can propose; humans gate what ships — same as any risky design doc.
• Greenfield AI-native infra might win in niches where nobody has legacy yet.
• Maximum efficiency is not maximum legibility — legibility pays rent for years.
• We already cannot hire enough people who understand Kubernetes; adding AI-specific layers hurts.
• Standards emerge from battle; let experiments happen in sandboxes, not core ledgers.