Harness Engineering: From Code Scaffolding to World-Model Stewardship

Abstract

The term harness in AI and software engineering has, in the span of fifty years, traveled an extraordinary arc — from "a collection of stubs and drivers configured to assist with the testing of an application or component" (Wikipedia, 2005) to a frontier proposition of AI control theory: that harness engineering defines, manages, and adjusts the world model the AI internally references. This paper surveys two decades of practice (2002–2026, with historical references back to the 1970s) across three lineages — software testing harness, robotics / embodied harness, and large-language-model agent harness — and argues that they converged in 2024–2026 into a single engineering discipline whose object is everything around the AI: tool mediation, context provisioning, policy enforcement, memory substrates, and most recently the AI's own model of the world. We organize the surveyed material along (i) a five-layer Jobs-to-be-Done taxonomy, (ii) a six-axis operational taxonomy, and (iii) a six-tension research agenda. We propose that the central methodological event of 2026 is the Stewardship Pivot — the shift from harness-as-controller-of-actions to harness-as-steward-of-internal-reality — and argue that this pivot reframes what AI engineering is in a way the field has not yet fully named. The survey draws on 489 unique sources across eight thematic clusters (60+ in each of software harness, robotics harness, and LLM agent harness; 50+ in world model concept, context engineering, multi-actor harness, formal verification, and harness ethics).

Keywords