From Harness Optimization to Recursive Self-Improvement: A Survey of Language Model Agents

Source abstract

Language-model agents rely on harnesses that govern tool use, context, memory, and execution. Understanding automated harness revision requires connecting improvement methods to the reusable functionality they produce. This survey examines three linked perspectives: improvement workflows, recursive scope, and concrete modifications across task domains. We compare search strategies, memory management, edit guidance, and integration with model training, and use an existing five-level framework to distinguish harness optimization from recursive self-improvement. Across coding, reasoning and information processing, interactive and embodied tasks, and visual tasks, we trace reported changes from starting capabilities to outcomes. Our synthesis identifies repairs at model–environment interfaces, cases where added tools and context fail to help, and conditions constraining recovery reuse. Cross-domain analogies motivate, rather than establish, transfer. We distinguish task-performance gains from improved subsequent optimization and identify priorities in reliable evaluation, attribution, transfer, and safety.