NeFut Logo NeFut
Admin Login

[CS.AI] Applying Anthropic Primitives at Large Enterprises: The Harness Paradigm for Knowledge Work

Published at: 2026-08-24 22:00 Last updated: 2026-08-29 12:04
#AI #LLM #Open Source

Frontier models have collapsed the cost of writing custom code: a niche problem that once required days or weeks of specialist effort can now be solved in an afternoon. By contrast, the cost of reviewing and maintaining that code has not dropped. Each solution drifts from the next, so understanding one often means reading its codebase from scratch. Large enterprises therefore build centrally governed solutions: at worst an off‑the‑shelf product, at best a graph‑orchestration framework or a low‑code platform wired for each use case. These approaches require custom development every time and have limited scope.

Enter the third option that escapes both constraints—the harness paradigm. Recent work treats the coding‑agent harness as enterprise infrastructure rather than a mere coding tool, converging on three findings:

  1. Harnesses are sufficient at the task level and outperform more elaborate architectures on enterprise workloads (see arXiv:2604.00073, arXiv:2604.13107).
  2. Harness choice accounts for most of the variance in agent benchmark results, more than model choice does (see arXiv:2605.23950).
  3. The gap between these findings and enterprise adoption is governance (see arXiv:2605.10223, arXiv:2605.18747).

We propose an architecture that closes that governance gap. A single unmodified harness runs as the backbone; the code remains identical across every deployment, so reviewing what gets built collapses to reading its instructions file. Section 4 describes four mechanisms that make this possible:

With this design, enterprises can keep the codebase uniform while achieving flexible permission management and auditability, dramatically lowering maintenance overhead and improving security.

Blogger's Review: The harness paradigm offers a pragmatic path for large organizations to balance scalability with governance, and it merits further experimentation in real‑world deployments.

Original Source: https://arxiv.org/abs/2608.20622

[h] Back to Home