Hermes Agent is an 🏷️#open-source autonomous AI agent framework from 📝Nous Research, released February 25, 2026 and built around a closed learning loop — it creates skills from experience, improves them during use, and grows more capable the longer it runs.
Hermes is not a coding copilot tethered to an IDE or a chatbot wrapper around a single API. It's a long-lived agent runtime designed to run on a $5 VPS, a GPU cluster, or serverless infrastructure with near-zero idle cost. After a complex task — typically five or more tool calls — Hermes autonomously generates a Markdown skill document following the agentskills.io open standard, capturing procedure, pitfalls, and verification steps. The next time a similar task appears it loads the skill and refines it rather than re-deriving the workflow from scratch. Nous Research calls this the closed learning loop: solve, document, retrieve, improve, repeat.
The architecture is provider-agnostic. A pluggable transport layer (AnthropicTransport, ChatCompletionsTransport, ResponsesApiTransport, BedrockTransport) routes inference through any compatible endpoint — Nous Portal, 📝OpenRouter, Anthropic, OpenAI, Google AI Studio, AWS Bedrock, xAI, MiMo, Kimi, 📝Ollama, or your own. Persistent memory uses SQLite FTS5 full-text search across every past session with LLM summarization for cross-session recall, paired with Honcho dialectic user modeling that builds an evolving model of who you are. Code and terminal execution run inside one of six hardened sandbox backends (local, Docker, SSH, Daytona, Singularity, Modal) with container hardening, namespace isolation, and a pre-execution scanner.
A single gateway process makes Hermes addressable from 20+ messaging platforms — Telegram, Discord, Slack, WhatsApp, Signal, iMessage, email, SMS, Matrix, Microsoft Teams, Google Chat, and the major Chinese platforms — so conversations stay continuous across devices and apps. Because it's built by a model-training lab, Hermes ships batch trajectory generation and Tinker-Atropos RL environments alongside the agent — the same toolchain Nous uses to fine-tune tool-calling models like its own Hermes-3 family (a 📝LLaMA 3.1 fine-tune). The result is an 📝AI framework positioned for compounding personal intelligence rather than ecosystem breadth — the direct architectural counterpoint to OpenClaw covered in 📝OpenClaw vs. Hermes.