A new model reveals that when large language model (LLM) agents retry tasks after failure, the previous failed attempts contaminate their context, leading to higher error rates. This research provides a framework for understanding and mitigating these failures, potentially improving the efficiency of LLM applications in multi-step processes.