Conceptual integrity and counting lines of code

In the latest episode of the Talking Postgres podcast, Claire Giordano and Simon Willison discussed the impact of AI on software development. Willison argued that there are times when lines of code can be a valid indicator of productivity. He explained that a skilled engineer could previously produce around 200 lines of high-quality, debugged code per day, while AI agents could potentially generate 1,000 lines of well-structured code in the same timeframe. This represents a significant increase in output.

However, Willison emphasized that while agents can accelerate development, they introduce challenges related to cognitive capacity and conceptual integrity. He used the analogy of the Winchester Mystery House to illustrate this point. Just as the house grew out of control due to the builder's psychic's influence, software built with agents can accumulate unnecessary features and "weird bumps" that compromise the software's overall integrity.

The key takeaway is that discipline remains crucial in software development, even with AI assistance. In the past, the time required to implement a feature discouraged engineers from pursuing low-value ideas. With agents, however, the rapid turnaround makes it tempting to add numerous features, leading to an increasingly complex and hard-to-manage codebase.

Ultimately, Willison argued that the new limiting factor is human cognitive capacity, necessitating a team of engineers to balance the workload and maintain the software's conceptual integrity.