Existing methods for evolving runtime harnesses for LLM agents treat each observed failure as evidence for modification. This can lead to model-specific accommodation, where the harness adapts to limitations of the underlying LLM rather than systematic harness deficiencies, impairing generalization. Ecdysis instead aggregates failure evidence across multiple task instances, promoting only recurring failure patterns into persistent harness changes. This biases evolution toward generalizable repairs.
Across multiple LLMs and benchmarks, Ecdysis improved the reasoning accuracy of evolved harnesses by 18.56% over existing methods while achieving up to 1.84x faster harness training. The resulting harnesses showed stronger cross-LLM generalization and lower inference-time token consumption, and the method enabled more data-efficient training. The key insight is that failures recurring across distinct tasks provide more reliable evidence for systematic harness issues than isolated failures.