TimeEvo: failure-driven tool synthesis lifts time series agent accuracy on every task and backbone

Jie Yang · hf · 2026-09-29

Time series agents answer analytical questions by calling tools, but tool libraries are usually hand-picked in advance. The authors identify two failure modes:

Human-agent tool misalignment: a library of 21 expert-curated tools helps some tasks and hurts others, dropping anomaly accuracy under every backbone tested;
Silent harm: one round of generic self-revision changed 147 answers and broke 56, while the final score moved by less than a point.

The root cause: whether a tool helps is decided question-by-question at runtime, yet tools are supplied in advance and judged by a single average.

TimeEvo clusters diagnosed failures into capability gaps, plans a measurement for each, synthesizes evidence-only tools to fill them, and admits candidates through a paired admission gate.

Experiments on ten time series QA tasks across three backbones show improvements on every task and backbone starting from an empty library, and a library grown on a cheap model still transfers to stronger ones. Code is open-sourced.

Original post →

TimeEvo: failure-driven tool synthesis lifts time series agent accuracy on every task and backbone

More from coding & agent