LLM-powered agents are increasingly applied to time-series problems, yet design choices vary widely across different tasks. This survey adopts a problem-driven taxonomy, organizing systems by the specific time-series challenges they address rather than by isolated technical components. We group existing work into four categories: forecasting and reasoning, augmentation and synthesis, anomaly detection and diagnosis, and decision support.
For each category we examine how task requirements shape the overall architecture, the way external tools are invoked, and the design of memory modules. For instance, forecasting tasks often employ sliding‑window memories combined with numerical computation tools, while anomaly detection relies more on historical pattern comparison memories and interpretability utilities.
We also compile representative datasets and experimental environments, and compare reported model performances under shared or closely related settings, enabling quick benchmarking. Our analysis reveals a lack of unified evaluation protocols for certain tasks and notable gaps in cross‑domain transfer and long‑term memory retention.
In summary, this survey offers a task‑oriented design guide for building LLM agents for time-series applications and highlights open challenges for future research.
Blogger's Review: The paper presents a clear structure and sensible categorization, providing practical insights for engineers building time-series LLM agents. It is a valuable read for anyone working in this area.