Nous Research's Hermes Agent Adds /learn Skill Builder and Claims Benchmark Wins

Nous Research has expanded its open‑source Hermes Agent with a new "/learn" command that lets the agent automatically gather material from local SDK folders, online documentation, recent conversation flows, or free‑form notes and generate a reusable skill file stored in ~/.hermes/skills/. This feature streamlines the creation of repeatable tools without requiring engineers to manually write skill descriptions.

In addition, Hermes Agent’s Mixture of Agents (MoA) presets—an approach that queries multiple language models and aggregates their responses—are reported to outperform Anthropic’s Claude Opus 4.8 and OpenAI’s GPT‑5.5 on benchmarks such as SWE‑bench Pro, achieving a 69.2% score versus GPT‑5.5’s 58.6%. The agent also includes persistent memory across sessions, a built‑in learning loop for skill creation, and tool integration for external APIs. While the full leaderboard has not been published, the company notes higher compute costs per query due to running multiple models simultaneously.