Time series agents answer analytical questions by calling external tools, but the set of tools they carry is chosen by humans before the agent ever runs. A new paper posted on Hugging Face, TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent, identifies two problems with that arrangement.
The first is Human-Agent Tool Misalignment: a library of 21 expert-curated tools helps on some tasks and hurts on others, dropping anomaly accuracy under every backbone the authors tested. The second is Silent Harm: a single round of generic self-revision changed 147 answers and broke 56 of them, while the final score moved by less than a point.
Both failures trace back to the same gap, according to the authors: whether a tool helps is decided question by question at runtime, yet tools are supplied in advance and judged by a single average.
To close that gap, the paper proposes TimeEvo, which clusters an agent's diagnosed failures into capability gaps, plans a measurement for each gap, synthesizes evidence-only tools to fill them, and admits the candidate library only through a paired admission gate.
Experiments across ten time series QA tasks and three backbones show that TimeEvo, starting from an empty library, improves accuracy on every task and every backbone. The authors also report that a library grown on a cheap model still gains when installed into stronger ones. Code is available on GitHub.