Agentic Workflows in Software Engineering

In software engineering, work is almost never purely mechanical.

The hard part isn’t typing code, it’s discovering constraints, modeling the domain, defining contracts, and making irreversible decisions explicit.

When that work is done, TDD / coding is simple, fast and reliable. Explaining intent to an AI, then reviewing and correcting its output, is times slower than writing the code as a sure-fire.

Agents help only after intent is frozen, when the work becomes repetitive and strictly bounded. They assist with some expansion and verification, not authorship.

Agents don’t replace engineers. They operate after 90% of engineering already happened, saving probably not more than 5-10% of engineering time.