AI accelerates code. Specifications accelerate the system.

Agentic leverage comes less from a spectacular prompt than from explicit decisions, constraints and verification.

Agents can generate a great deal of code very quickly. That does not replace business judgement or verification. It makes both disciplines more important.

The prototype-that-ships trap

A result that appears to work is not yet a system. Exceptions, real data, responsibilities and security rules usually arrive after the first demo.

Reverse SDLC

Starting with work situations, making decisions explicit, modelling constraints and testing a prototype reduces the cost of wrong code before it exists.

The developer role changes

Useful speed comes from defining intent, evaluating output and installing guardrails. Agents implement; humans own the trade-offs.