Four Essays

AI made speed free. You can produce code, configs, documentation at near-zero cost. But it didn't make good code free. The bottleneck shifted from "can we build it fast enough" to "can we build it well enough." Quality is the scarce resource. Tooling is how you produce it.

Quality requires refinement. Refinement produces learning. Learning should be encapsulated. When AI builds your systems, the code is a side effect — the process that produced it is the real deliverable.

The default agent escape hatch is bash — god mode with no schema, no guardrails, no observability. The fix: identify every shell command the agent reaches for, then build a typed tool that does the same thing better.

The direct intellectual argument. Yaron Minsky's make illegal states unrepresentable maps precisely to AI tool systems. Type checking is scope enforcement. Unsafe blocks are flow suspension. The trust boundary question is the one programming languages have been answering for thirty years.