Claude Code’s father: Harness has only a six-month shelf life, let it off the leash

Boris Cherny said Claude Code was heavily simplified by removing prompts and harness code, arguing developers should not predefine model needs but refine systems through deletion, testing, and iteration.

He stressed that model capability often runs ahead of product packaging, using the ideas of Product Overhang and Unhobbling to argue that AI products should unlock the model’s full potential.

Anthropic’s approach pushes models to self-verify and work independently for longer periods, moving away from tightly controlled, fixed-prompt product design.

The interview signals a shift in frontier AI product building toward empirical, experiment-driven design with less scaffolding, especially for coding tools and agentic products.