Language-model agents increasingly handle complex tasks by gathering information in parallel and asynchronously. A new preprint from arXiv examines how the timing of that information delivery affects the agents' search behavior. The authors argue that the scheduler—the component that decides when evidence arrives—acts as a hidden retriever, shaping what the agent ultimately sees and uses.
The study uses controlled replay experiments across two model families and three benchmarks. By holding the search process constant and varying only delivery timing, the researchers isolate the effect of asynchrony on evidence acquisition. The abstract reports that delivery timing redirects search, suggesting that when results arrive can be as important as which results are retrieved.
This finding has practical implications for designing search agents: optimizing the scheduler may be a low-cost way to improve agent performance without changing the underlying retriever. However, the abstract does not specify whether the effect is positive or negative, nor does it detail the magnitude. Further reading of the full paper would be needed to understand the direction and practical guidance.