Agents made new features so cheap that code loses conceptual integrity

AI agents are genuinely changing developer metrics

In a recent Talking Postgres podcast episode, Simon Willison opened a debate about productivity and the limits of AI agents in software development. He notes that while agents are often criticized for generating thousands of lines of low-quality code, they genuinely change the metrics for experienced engineers. Increasing daily output from 50 to 1000 debugged lines is achievable, but it hits human cognitive limits.

Cheap code builds chaotic software labyrinths

The fundamental problem Willison describes uses the term conceptual integrity from the classic book The Mythical Man-Month. Good software has clear boundaries and logic. But with AI agents, adding a new feature is a matter of one prompt and five minutes. The result is systems that grow uncontrollably in all directions, resembling bizarre houses built over decades without a blueprint.

Time used to act as a natural filter for nonsense

In the pre-agent era, the sheer time required stopped developers from adding unnecessary features. An idea that would take a week of work was automatically scrapped. Today, when the same experiment takes an hour, it is much harder to say no. Tools have made experimentation cheap, but paradoxically removed the friction that kept software clean.

Approving pull requests will demand more discipline

If creating code costs almost no time, checking its viability shifts from the planning phase to the merge request phase. Teams will have to be much stricter about whether a new feature makes sense at all, because the complexity of implementation no longer works as an effective filter.

Lilith's verdict

Technical debt no longer comes from being too lazy to rewrite something. It now comes from it being too easy to add something.