Ecdysis evolves LLM-agent harnesses only from failures that recur across tasks

Self-evolving runtime harnesses can substantially improve LLM agents, but existing failure-driven approaches treat each observed failure as direct evidence for harness modification. A key problem: a failure can reflect either limitations of the underlying model or systematic deficiencies of the harness, and optimizing against individual failures induces model-specific accommodation that impairs generalization across tasks and models.

Ecdysis's insight is that failures recurring across distinct tasks provide stronger inductive evidence for systematic harness deficiencies than isolated failures. It aggregates failure evidence across task instances before promoting recurring failure patterns into persistent harness evolution, and employs collaborative failure analysis to refine modification specifications.

Across multiple LLMs and benchmarks, Ecdysis improves the reasoning accuracy of evolved harnesses by 18.56% over existing harness evolution while achieving up to 1.84x faster harness training, with more data-efficient training. Fine-grained analysis shows it reduces model-specific accommodation during evolution, and the resulting harnesses exhibit stronger cross-LLM generalization and lower inference-time token consumption.

Failures recurring across distinct tasks are stronger inductive evidence of systematic harness deficiencies than isolated ones.

Aggregating first keeps one model's quirks from being baked into the harness.

Collaborative failure analysis refines modification specifications before edits land.

Cross-LLM evaluation shows the edits generalize instead of accommodating a single model.

Before patching a system around a failure, ask how many independent contexts produced it. Recurrence across contexts separates defects of the scaffolding from limits of the underlying worker.

Posted as a v1 preprint on 10 September 2026 (v2 on 20 September); reported across multiple LLMs and benchmarks with fine-grained analysis of reduced model-specific accommodation. No production deployment reported.