Thin Harness, Strong Contracts: Production-Oriented Agent Harnesses for Stateful AI Agents

Preprint artifact package for Thin Harness, Strong Contracts: Production-Oriented Agent Harnesses for Stateful AI Agents. The package includes the checked manuscript PDF, LaTeX source, bibliography, a 24-case replayable pilot benchmark, live DeepSeek and Kimi raw run logs, regenerated result tables, uncertainty tables, figures, evidence snapshots, reproducibility scripts, and release documentation. The paper argues for production-oriented agent harnesses that stay thin around model cognition while enforcing strong contracts at system boundaries: tool calls, permissions, state diffs, replay, memory scoping, audit, and human review.