Agent Execution Harnesses: Architecture, Constraints, Sovereign Variants, and Collective Dynamics — Volume I

This paper synthesises the emerging literature and engineering practice surrounding the agent execution harness—the runtime control layer that transforms a foundation model into a reliable agent. We formalise the core equation Agent = Model + Harness and examine how harnesses are constituted across observation, context, control, action, state, and verification/governance layers. Drawing on the ReAct framework and the control-plane versus intelligence-plane distinction, we analyse dataset construction for major benchmarks (SWE-bench, GAIA, WebArena, OSWorld), multi-layer validation protocols, and the distinctive requirements of sovereign AI harnesses. We then detail the technical, economic, and jurisdictional constraints harnesses face, the mechanisms by which they unlock latent model capabilities while mitigating context rot, tool overload, and non-determinism, and the architectural relationship between individual harnesses and multi-agent coordination that can be analogised to topological braids. The paper concludes that, for long-horizon tasks with comparable frontier models, harness design is frequently the binding constraint on reliability, auditability, and sovereignty.