Microsoft Research (with University of Chinese Academy of Sciences) released STRACE, a framework that treats agent execution logs as causal dependency graphs rather than linear text, enabling automated diagnosis and prompt-policy optimization for long-horizon AI agents. On a formal verification benchmark (VeruSAGE-Bench), STRACE improved success rates from 42.5% to 58.5% (absolute +16 points), outperforming all tested baselines including TextGrad and GEPA; gains on WebArena reached +12.9 points overall. The core mechanism filters redundant failure traces, then uses backward dependency-graph slicing to isolate root-cause modules rather than downstream symptoms, injecting corrective heuristics only into the implicated prompt policy. For financial institutions evaluating agentic AI deployments, this class of automated agent-optimization tooling directly bears on the practical reliability gap between single-shot model capability and multi-step autonomous task completion.
293 other tracked developments touching the same claims.