AI Agents Speed Up Coding but Weaken Consistency

Simon advocates for a pragmatic reading of code lines in the era of AI agents. He argues that these tools can accelerate writing to the point of delivering a feature in minutes, but that overall coherence suffers if discipline wanes. For him, the true limit becomes cognitive capacity, justifying the maintenance of teams.

Agents deliver a function in 5 minutes, and coherence wavers

With coding agents, a feature idea can be realized in five minutes after a simple prompt, according to Simon. He describes software that develops in small spurts in unexpected directions. For him, this ease of continuous addition, made possible by a significantly lower cost, leads to the collapse of conceptual integrity and complicates decision-making.
The concept of conceptual integrity, presented in The Mythical Man-Month, refers to software without surprises, with a just scope, and whose parts fit together coherently. Simon believes that when the time required no longer naturally imposes a filter, discipline weakens: an idea abandoned if it were to take a week becomes more acceptable if it only takes an hour.
Claire Giordano invokes the analogy of the Winchester Mystery House. Simon reports that it has 140 rooms, and that its owner, the widow of the inventor of the Winchester rifle, supposedly built it relentlessly for 40 years on the advice of a medium. However, he notes that credible sources cited by Wikipedia contest this story of the medium.

Productivity: 50–60 lines/day, 200 a peak; 1,000 with agents

Many consider it inappropriate to measure productivity in lines of code. Simon disagrees and argues that there is a strict limit that makes this indicator usable. He recalls that in the past, an engineer could produce a few hundred lines ready for production in a day.
He estimates a very productive day at 200 lines of functional and debugged code, whereas, most of the time, 50 or 60 lines were completed. He adds that agents can allow reaching 1,000 debugged lines, provided the same level of quality (maintainability, testing) is maintained. According to him, this threshold requires a lot of skill, knowledge, and experience, typical of senior engineers.

The limit becomes cognitive capacity and necessitates teams

Simon claims he can accomplish much more solo with agents than without, to the point of questioning why a company would keep multiple engineers. He dismisses this hypothesis, reminding that a team of one person is very poorly designed and that the new constraint lies in cognitive capacity.
He argues that he can produce code one hundred times faster, without being able to manage one hundred times more code. In his view, teams remain necessary to balance cognitive load. These positions were presented during an episode of Talking Postgres recorded last week with Claire Giordano, from which the main points are derived from a slightly edited transcript.