Are AI coding agents actually changing software economics, or only development speed?

The strongest claim around coding agents is that software becomes dramatically cheaper because implementation time collapses. I think that mixes up two different costs.

Writing code is one part of software production. Deciding what to build, integrating it with old systems, verifying behavior, operating it, supporting users and accepting liability are different costs. Agents clearly compress some implementation work. They may also increase the amount of software organizations attempt to maintain.

So the economic question is not simply whether code becomes cheap. It is whether the total cost of reliable software falls after organizations respond to cheaper code by producing more of it. My expectation is that prototypes become radically cheaper while dependable production systems become somewhat cheaper but more numerous.

The maintenance multiplier is the part I see being ignored. I can create internal tools I would never have justified building a few years ago. That is a productivity gain, but it also means I own more systems.

Cheap implementation also changes what gets tested. Teams can afford more failed experiments. Economically that may matter more than shaving a fixed percentage from the cost of a successful product.

Security review is a useful counterexample. More generated code means more surface area. If review capacity does not scale at the same rate, the bottleneck simply moves.