LongHorizon-Harness is an open-source 'loop engineering' system from Alibaba's AMAP-ML research group that wraps existing coding and computer-use agents -- Claude Code, Codex CLI, OpenCode, and DeepSeek Harness -- in a durable execution loop so they can keep working toward a single goal across desktop apps and the terminal for extended sessions instead of stalling after one exchange. Rather than training a new model, the harness assigns three role boundaries around whichever backend agent is plugged in: a Manager that reconstructs the remaining work each round from the original goal and verified progress, an Executor that carries out one bounded step with a fresh context, and an Auditor that independently inspects files, UI state, logs, and tests rather than trusting the Executor's own report. Only auditor-verified results are checkpointed as trusted state; rejected attempts are kept as evidence and fed into the next planning round instead of being discarded. The project ships a one-command CLI installer, a browser-based Dashboard (React/FastAPI) for starting runs, picking a model per role, approving actions, and sending mid-run instructions, plus a pluggable AgentAdapter and Environment protocol so any model or backend can be substituted. It reports measured completion-rate gains on WeaveBench, OSWorld 2.0, and Terminal-Bench 2.1 using the same underlying model and execution backend, isolating the harness itself as the source of improvement. The project has an accompanying arXiv paper and reached #1 on the Hugging Face Daily Papers weekly ranking in August 2026.