Probabilistic engineering and the 24-7 employee

Software is moving from deterministic engineering to probabilistic engineering. In AI-native teams, code is increasingly something people believe works rather than know works, and the bottleneck shifts from typing to direction, selection, and validation. The essay argues that the teams already using agents at scale are shipping much faster, but also taking on new risks around review quality, role fragmentation, and training.

Reading notes

#

• Code generation is getting cheap, but validation is not, so review becomes the scarce and difficult part of the workflow.
• Inside AI-native teams, some people move upward into higher-leverage roles, while others are pushed into spec writing, review, and agent babysitting.
• Jevons’ paradox is used to argue that cheaper code leads to more software being built and shipped, not less.
• The author describes a shift from deterministic engineering, where failures are reproducible, to probabilistic engineering, where confidence in correctness is only partial.
• Frontier teams already rely on agents that open pull requests, review work, self-heal test suites, run experiments, and update documentation.
• Different domains will adopt this shift at different speeds, with regulated systems staying deterministic much longer than consumer software and internal tools.
• The “agentic fleet” metaphor describes coordinated agents working overnight under human command, with morning triage and redirection.
• Teams are urged to build for the model they do not have yet, since the current model is framed as the weakest they will use.
• The essay warns that juniors who rely on AI too early may never develop the craft and judgment that come from doing hard debugging and evaluation themselves.
• The closing point is that organizations need both agentic speed and human skill, because the 24-7 employee is a wager on human judgment staying sharp enough to supervise the system.