← All benchmarks

Hermes Bench

Agents · 2026-10-06

Nous Research's suite of 150 agent tasks in 25 categories, mostly Hermes skills and research work plus diagrams and art, memory, browser tool use and one Git-safety task. Each task gives the agent a seeded workspace and a request, sometimes with scripted follow-up turns, then grades the files and state left behind with deterministic checks, an LLM judge or a vision judge. The score is mean reward with partial credit, one pass@1 run per task in the Hermes Agent harness.

Top models (higher is better)