Tech News,
The Latest First.

Why Your AI Agent Keeps Failing on Opus 5.5 Is the Harness, Not the Model

A guide doing the rounds among developers makes a claim that cuts against most people's instincts: the reason your AI agent keeps failing on Claude Opus 5.5 is almost never the model. It is everything wrapped around it, a layer the author calls the harness. The argument is backed by a clean before-and-after test on 20 real tasks, where the numbers moved sharply without the model or the prompt changing at all. For anyone who has watched an agent say "done" over a red test suite, it lands hard.