Harness Paradigm Paper Argues One Unmodified Coding Harness Should Be Enterprise Infrastructure, Not a Coding Tool
Applying Anthropic Primitives at Large Enterprises (arXiv 2608.20622, Aug 20) cites converging results that harnesses suffice at task level and outperform more elaborate architectures on enterprise work, and that harness choice accounts for more variance in agent benchmark results than model choice does. Its proposal is one harness running unmodified as the backbone so that reviewing what gets built collapses to reading its instructions file, with four mechanisms: credential-scoped tooling giving each backend one generic request tool plus a scoped credential, authorization logic living outside the harness, and registration as a side effect of pushing code. The reference implementation is built on microcc, published on PyPI.
↳ Follow the thread