Harness engineering is a discipline that aims to make AI coding agents reliable in production. However, it has been found to be insufficient in ensuring the success of software factories. The primary reason for this is that harness engineering focuses on the environment around AI models, but fails to address the fundamental issue of model training and codebase maintainability [1]. According to Dex Horthy, founder of HumanLayer, 'the harness is not enough' [2].
The narrative that 'you are the bottleneck' and that 'just shipping more' is the solution has been debunked [3]. In fact, the cost of bad architecture is measured in weeks, months, or even years, making it difficult to optimize for [4]. Furthermore, models are good at some things but not others, and optimizing systems within these constraints is crucial [5]. To address these issues, it is recommended to turn the lights back on and plan up front, including product review, system architecture, program design, and vertical slices [6].
This approach can help move 2-3x faster while maintaining a human level of code quality. In conclusion, harness engineering is not enough to ensure the success of software factories, and a more comprehensive approach that addresses model training and codebase maintainability is necessary.
