The Bottleneck Moved: Why Cheap Code Still Makes Expensive Software

coding agents

The question

Two developers open the same agent, on the same model, on the same Monday.

By Friday, one has shipped a refund feature that survived production. The other has 4,000 lines of tidy-looking code, a red CI pipeline, and a growing suspicion that the agent “understood” a different payment system than the one the company runs.

The usual explanation is “better prompting.” I think that explanation is mostly wrong. Prompting is the visible part. The difference sits underneath it: what the agent knew before the first message, what was allowed to count as “done,” and who was holding the pen when the definition of “correct” got written.

This letter takes that apart. The claim is narrow: when code generation gets cheap, the scarce resource becomes everything that makes code trustworthy. Specs, invariants, boundaries, and independent verification used to be implicit in how a team worked. With agents, they have to become explicit artifacts.

Part 1: What this actually is

Strip away the buzzwords. Three different things get blurred together under “AI productivity”:

• Code production: generating code, editing files, making commits.
• Software delivery: code that has been integrated, tested, reviewed, and deployed into something that runs.