From Question Answering to Task Completion: A Survey on Agent System and Harness Design

This survey presents an analysis of LLM-based agents, emphasizing their evolution from passive question answering to active task completion through the integration of execution harnesses. It decomposes agent systems into six runtime responsibilities—observation, context, control, action, state, and verification—and explores how these interact with foundational models, highlighting the importance of harness configurations on task efficiency and reliability. The findings underscore the need for a holistic approach to agent design that considers model-harness co-evolution, open challenges in evaluation, and safety, which are critical for practitioners developing advanced AI systems.