Benchmarking Hermes Agent: PAST-Bench, FinSkillBench, HarnessRisk, and SWE-bench Pro

PAST-Bench reports recursive self-improvement gains for Hermes+, while FinSkillBench evaluates Hermes across 8 models and 5,280 investment-management episodes. HarnessRisk adds lifecycle-oriented comparisons with OpenClaw and Nanobot, and NVIDIA NeMo Gym now documents a structured SWE-bench Pro integration for Hermes, although the reviewed integration provides no performance results. Long-horizon memory and reliability remain unresolved.