Rohan Paul

If an agent rewrites its runtime for every bad decision, it can learn the wrong lesson Fix recurring failures across tasks, not every failure. Agent runtimes should learn from failure patterns, not failure counts: ECDYSIS grouped recurring problems across tasks and delivered higher accuracy, faster training, and stronger cross-model transfer. Patch every miss, and you can accidentally hard-code one model's bad habits into the system. ECDYSIS instead looks for the same kind of failure across different tasks before deciding the harness itself needs fixing. – arxiv. org/abs/2609.11677 Title: "Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents"