Marty Lamb argues teams still own AI-written code they do not understand

This is an essay rather than a news item, and it is short. Marty Lamb published it on 25 August. The argument is worth the ten minutes.

Two tasks that are now separate

Writing software has always included a second task: understanding the software. Not designing it so others can follow it, which is a separate discipline with its own literature, but understanding your own code, the code you typed yourself.

That understanding always took effort. It happened automatically, because writing something forces you to reason about it. Nobody has needed to think of it as a separate cost, because it never took separate time.

Lamb's point is that agents have now separated the two. Generating code now costs almost nothing, understanding it still takes effort, and it is now possible to release a system nobody on the team has a model of. Have the agents review and test it as well, and the mental model becomes fully optional.

The part that does not become optional

When you build software you own it, and not only in the copyright sense. You take on quality, security, debugging, operation, maintenance and change over time.

None of those get cheaper because the code was generated. Every one of them is easier with a mental model of the system and much harder without one.

So his reframing of the decision: the question is no longer whether you can build a thing, it is whether you want to own it. He is direct that this should be a conscious choice at the start, not a responsibility discovered after the code already exists.

He is honest about the exception

For sufficiently low-stakes software, he says, generate as much as you like, and describes doing exactly that himself for personal tooling and small projects that he would otherwise never have started.

That matters, because the argument is easy to misread as being against generated code. It is not. It is about matching how much understanding you invest to how much the thing costs you when it breaks.

Where he thinks this leads

He calls human understanding of generated code the new slowest step.

His open question is what software would look like if keeping it understandable to people were a main design goal. Not understanding every line at all times, but pieces that can be understood individually, with boundaries clear enough to look at the whole system or at one part and reason safely at each level.

He says he is still developing those ideas, so there is no framework at the end of this. The new way of framing the question is the main result, and it is a good one to share with a team before the next build starts.

One line to take into the next planning meeting: AI can produce complexity in large amounts.

Source

• AI-Written Code Is Still Your Code. Are You OK With That?, Marty Lamb, 25 August 2026