Harness-Native Software Engineering: The Control Plane of Coding Agents
Abstract
Coding-agent systems are increasingly shaped by infrastructure outside the base model: repository guidance, tool permissions, execution sandboxes, persistent memory, verification loops, recovery mechanisms, and delegation logic. For long-horizon software tasks, those design choices often determine what an agent can see, what it may do, and whether it can recover from failure.
This paper introduces Harness-Native Software Engineering as a framework for analyzing that shift and formalizes the surrounding runtime layer as the Agent Harness Control Plane (AHCP), with eight functions: context ingress, action mediation, execution substrate, state persistence, verification and review, recovery and debugging, delegation and coordination, and governance and audit.
Drawing on recent system documentation, benchmark papers, configuration studies, adoption analyses, and security research, the paper advances five claims. Coding-agent performance is harness-sensitive. Recoverability matters alongside one-shot success. The effective security boundary sits at the harness interfaces through which tools, protocols, permissions, and sandboxes are exposed. Teams are externalizing agent behavior into version-controlled artifacts such as manifests, skills, and custom-agent definitions. And benchmark rankings can change when harness conditions or task realism change.
The contribution is conceptual and methodological: a reusable unit of analysis for coding agents, a comparative matrix of representative systems, a reporting instrument called the Harness Condition Sheet, a threat model for the harness layer, and a harness-sensitive protocol for future empirical work.