Researchers propose an architecture for large enterprises to use the 'harness paradigm' as a way to manage and govern AI systems. This approach treats the coding-agent harness as enterprise infrastructure, rather than just a tool, and converges on three key findings: harnesses outperform more complex architectures in enterprise work, the choice of harness is crucial, and governance is the main barrier to adoption. The proposed architecture closes this gap by allowing one unmodified harness to run as the backbone, with code identical across deployments, making it easier to review and maintain. --- Why it matters: This matters because large enterprises struggle to manage and govern AI systems due to the complexity of custom code and limited scope of off-the-shelf products. The proposed architecture offers a scalable solution that can be used across multiple use cases, making it an important development for researchers and engineers working on enterprise AI projects. Source: https://arxiv.org/abs/2608.20622