From Prompt to Harness: An Engineering Methodology for AI Agents-01

Preface Over the past few years, the most common question around artificial intelligence—and the one most likely to provoke strong emotions—has been this: will models replace programmers? It is, of course, a compelling question, because it speaks directly to professional anxiety, industrial narratives, and our imagination of the age. But it also has an obvious flaw: it focuses too early on whether people will be replaced, while overlooking how work itself is being reorganized. The deeper transformation usually begins with the latter. Hold two scenes in your mind side by side. In the first, OpenAI used Codex to build internal products within a matter of months under almost extreme constraints: no one hand-wrote application code, and code, tests, documentation, CI, and observability were all produced by agents. In the second, METR invited senior open-source developers who were deeply familiar with their repositories to use early-2025 AI tools in real projects. The result was not faster d