Nous Research's new agent creates its own skills as you chat

Nous Research just open-sourced Hermes Agent, an autonomous assistant designed to grow smarter the more you use it. Instead of treating every chat as a blank slate, it comes with a built-in learning loop that extracts completed workflows and turns them into reusable skills.

Most current agent frameworks rely entirely on giant context windows or rigid prompts. Hermes Agent takes a different path: after completing a complex job, it autonomously writes a skill, updates it during subsequent runs, and uses FTS5 search with dialectic user modeling to remember your habits across sessions.

What it does

You can interact with Hermes Agent through a terminal UI or let it run headlessly on a remote server. A single messaging gateway connects directly to Telegram, Discord, Slack, WhatsApp, and Signal. That means you can leave it running on a remote machine, send a voice memo via Telegram, and let it carry out tasks in the background.

It handles unattended automation right out of the box. Through an integrated natural language cron scheduler, you can instruct it to run daily market reports or perform weekly audits and post the summaries straight to your chat channels.

For parallel heavy lifting, it can spawn isolated subagents or collapse complex multi-turn pipelines into zero-context Python RPC scripts. If you are migrating from tools like OpenClaw, Claude Code, or Cursor, it includes built-in importers to bring over your personas, memories, and existing skill directories.

How it runs

Hermes Agent supports virtually any backend provider without lock-in. You can hook it up to OpenAI, OpenRouter, local endpoints, or Nous Portal with a single command.

Infrastructure costs are remarkably low. It works on a five-dollar cloud VPS or serverless setups like Daytona and Modal, where the environment hibernates when idle and costs virtually nothing between runs. Cross-platform support covers Linux, macOS, native Windows without requiring WSL, and even Android via Termux.

How to get started

The code is fully open-sourced under the MIT license on GitHub under NousResearch/hermes-agent. Linux and macOS users can install it with a single bash curl command, while Windows users can install via PowerShell. Once set up, running the hermes command drops you into the terminal environment.

Why it matters

We think Hermes Agent takes the right practical approach to agent design. Instead of confining assistants to a web window or demanding expensive dedicated hardware, it meets users inside their existing chat apps and runs cheaply on lightweight cloud boxes.

If you want an assistant that handles recurring workflows and remembers preferences across tools, it is well worth exploring today. The biggest challenge will be keeping an eye on permission boundaries: letting an agent autonomously write skills and call local tools across messaging platforms means you will want to keep approval whitelists strictly configured.