Study Finds Self-Improving AI Agents Fragile Due to Noise and Task Order

A new arXiv preprint (2608.18066v1) re-evaluates memory-based self-improving agents—systems that learn from an online stream of tasks and maintain a textual memory bank. The study broadens evaluation along two axes: multiple runs to quantify variance and random shuffling of tasks to test task-order effects. Findings reveal that agent evaluation is inherently noisy in complex environments and multi-step tasks, and the self-improving loop can amplify that noise. Additionally, improvement depends heavily on task order; prior works often used default sequences that impose an implicit curriculum. The paper highlights the overlooked reliability issues in these methods and calls attention to underspecification in current evaluation protocols.