The next bottleneck for AI in software development is whether your code is optimized for AI understanding.
We may be creating legacy code faster than ever.
How legacy systems used to form
Historically, legacy systems became difficult to change over decades.
A monolith grew. New functionality was layered onto old decisions. Architectural boundaries blurred. Business logic accumulated in unexpected places. The people who understood why things worked that way eventually left.
How AI compresses the legacy cycle
AI may compress that entire cycle.
Coding agents make it possible to produce significantly more code, much faster. But more code is not necessarily more understandable software.
If teams optimize primarily for how quickly an agent can generate a working change, they may be overlooking a harder question: Will the next agent be able to understand that code correctly?
Why today’s AI-generated code could become tomorrow’s legacy
Today, we see the problem most clearly in older systems. Agents struggle with architectures that have accumulated years of implicit decisions, hidden dependencies, and inconsistent patterns.
But soon, we may see the same problem in code written today—with AI.
Not because the code is old, but because it was generated faster than the underlying architecture could remain coherent.
The agent that creates a feature today is not necessarily the agent that will maintain it tomorrow. It may have a different context window, different instructions, different tools, and no access to the reasoning behind the original implementation.
A new standard: code optimized for AI understanding
That suggests a new standard for software quality.
Code should not only run correctly.
It should be:
- Easy for an AI agent to understand.
- Hard for an AI agent to misunderstand.
- And economically viable for an agent to explore, reason about, and safely change.
Rethinking how we measure AI productivity
This changes how we should think about AI productivity.
Lines of code produced, tickets completed, and time saved tell us what happened today. They tell us very little about the cost of the next change. The real test of AI-generated software is not how fast an agent can write it. It is how reliably the next agent can work on it.
The legacy problem is not going away.
We may simply be accelerating it.
The real test of AI-generated software is how reliably the next agent can work on it.
