Agentic software development is starting to rhyme with a wave we already lived through: classic software engineering.
A simple greenfield function is uncomplicated to build and not particularly difficult to test. But an agentic workflow with multiple agents? That looks a lot like refactoring, or like building a complex enterprise system. Suddenly the old disciplines come rushing back. We have to think about overloading the architect with too much to evaluate in detailed design. We have to think hard about dependencies. We need documented system files, requirements, and evaluations, so development is a repeatable act, not a one-time one… so we can pick up where we left off.
Flooding a context window with information is not totally unlike flooding a human brain with too much to digest. A powerful frontier model can hold far more context than a human brain. Mine included. But we expect a lot more from these models, and we throw a ton more at them than we ever would at a person. So we run into the constraint again quickly. Right?
We’re not reinventing software engineering for agents. We’re rediscovering why it existed.
A question worth taking into Monday: could your most complex agentic workflow be rebuilt, or handed off, from what’s written down today — the requirements, the dependency map, the evaluations? If the honest answer is no… I’d start there.
Originally posted on LinkedIn.
RationalAI helps organizations turn ideas like this into results: AI strategy and advisory services. Previous article: The Confidence Transfer: When the Room Trusts the Machine Over the Expert
