STRACE framework mines failure patterns and extracts root causes for agent optimization

Researchers introduce STRACE (Structural TRajectory Analysis and Causal Extraction), a framework designed to optimize long-horizon agents by constructing high signal-to-noise contexts from execution traces. The method addresses the inefficiency of redundant and heterogeneous data by mining failure patterns at the batch level to filter out redundancy.

• At the batch level, STRACE mines failure patterns to filter redundant traces and retain representative failures.
• Within selected traces, it performs causal localization over a textual dependency graph to remove non-causal steps and identify true root-cause modules.
• On the VeruSAGE-Bench formal verification task, STRACE improved success rates from 42.5% to 58.5%, a 1.4x improvement over human-expert designed agents.

The framework significantly outperforms standard context-filtering baselines by providing precise optimization signals without discarding causally important evidence.