A new research agent fixes its own tool kit by studying its mistakes, instead of relying on tools engineers picked in advance.
Researchers describe TimeEvo, a time series analysis agent that diagnoses its own failures and builds new tools to close the gaps. The team first identifies two problems with today's approach: a curated library of 21 expert tools boosts some questions but drops anomaly-detection accuracy on others, no matter which underlying model runs it. Worse, a single round of standard self-revision changed 147 answers and broke 56 of them, while the overall score barely moved. TimeEvo responds by clustering failures into capability gaps, designing a measurement for each gap, generating evidence-based tools to fill it, and only admitting a new tool through a paired test that checks it actually helps.
The findings expose a blind spot in how these agents get built: usefulness gets decided by an average score, not by what a specific question needs, so a tool can quietly hurt as often as it helps. It also shows that self-revision, an increasingly common patch for agent errors, can look neutral in aggregate while scrambling scores of individual answers underneath.
Starting from an empty tool library and testing across ten tasks and three model backbones, TimeEvo improved accuracy everywhere, and a library trained on a cheap model still worked when moved to a stronger one - a cost-saving detail worth watching as more teams try to make agents fix themselves rather than hand-tune them forever.