SeisAgentBench: A Failure-Aware Core Benchmark for Multi-Agent Earthquake Rupture Inversion and Forward Validation

Earthquake risk-analysis workflows now span rupture inversion, wave propagation, hazard estimation, structural response, and regional impact modeling. We present SeisAgent, a dataset proposal for evaluating multi-agent orchestration across Grond, WISP, OpenSWPC, SeisSol, SPECFEM, SW4, OpenSHA, OpenQuake, ShakeMap, gmprocess, ObsPy, Pyrocko, NHERI SimCenter EE‑UQ/R2D, and a GeoGPT‑based evaluator agent. Language agents coordinate deterministic engines while a shared schema preserves episode state, provenance, governance, and cross-backend comparability.