METR study finds AI slows experienced developers

wpcms.computer.org4 min read

Drawing on operational experience with an internal multi-agent development platform, the author argues that the unit of work making agentic software development tractable is the specification rather than the prompt. By shifting to an approach where no code is generated without an approved specification stored alongside the repository, teams can eliminate prompt-based ambiguity, enforce architecture through structured templates, and use adversarial AI reviewer loops to catch errors early. Open article →