Quick vibes check: a 78-year-old marketing exec who had never written code shipped a working robotics app this week.
That stat comes from Clement Delangue’s thread on Hugging Face’s Reachy Mini app store, which crossed 300+ live apps and roughly 10,000 robots deployed worldwide.
The exec used natural language to build it without any Python or special robot software; in the time it would have taken you to install ROS, this person built a thing the robot now does.
These stories show a pattern
AI is becoming a tool that lets people build things they couldn’t before. Anthropic’s new Dreaming feature lets agents process past sessions overnight and write new memories for themselves while you sleep. And Nous Research’s Hermes Agent shipped a major release this week, pushing the same idea further with a persistent personal agent that learns your specific work over time.
Speaking of Hermes
If you don’t know, Hermes is kinda like the successor to OpenClaw (the personal AI assistant that defined this category over Christmas).
Yesterday, it shipped v0.13.0 “The Tenacity Release”, with 864 commits from 295 contributors in one week and 8 critical security holes closed. (One was a Discord bug that let bots message users across servers they shouldn’t reach.) About 30% of OpenClaw users have switched per Reddit sentiment surveys, citing easier setup, better memory defaults, and a self-improving learning loop.
Here’s what happened
Hermes Agent launched in February 2026 from Nous Research, the lab behind the Hermes model family. 135,000+ GitHub stars, MIT licensed, ships with 40+ bundled skills (modular instruction packs the agent reuses).
The architecture is built around a closed learning loop. After a complex task, Hermes enters a “Reflective Phase”: it analyzes what worked, extracts reusable patterns, and writes a new skill file encoding the solution. Next time a similar task arrives, it queries its own skill library instead of reasoning from scratch.
Three memory layers (session, episodic via SQLite, procedural skills). Runs on a $5 VPS, GPU cluster, or serverless. Model-agnostic; works with OpenRouter, Anthropic, OpenAI, Nous Portal, Kimi, MiniMax, GLM, or your own endpoint.
Talk to it via Telegram, Discord, Slack, WhatsApp, Signal, Email, or CLI. Yesterday’s release added Google Chat as the 20th platform, plus durable multi-agent Kanban with heartbeat, zombie-worker reclaim, retry budgets, and a hallucination gate. It also added persistent /goal for long-running tasks, post-write file linting on every edit, and session auto-resume when the gateway restarts mid-task.
Installation is a one-line curl installer that auto-handles all dependencies (Python 3.11, Node.js, ripgrep, ffmpeg). Run hermes setup and the wizard auto-detects ~/.openclaw, offering to import settings, memories, skills, and API keys (ask your regular AI chatbot to help you set it up if that’s confusing)
Why this matters
OpenClaw built the category by organizing everything around a messaging hub; Hermes flipped the design, putting the agent’s learning loop at the center. Both agents can have AI-written skills, but Hermes’s loop is automatic.
OpenClaw skills are runbooks that you (or an AI you prompt) write up front. Hermes pauses every ~15 tool calls and after complex tasks, reflects on what just worked, writes a Markdown skill file capturing the pattern, then refines it the next time. Compounding is built in.
Our take
Hermes isn’t strictly better. OpenClaw still has 24+ messaging integrations (vs. Hermes’s six), more security scrutiny, and transparent file-per-memory you can inspect. Many power users run both, with OpenClaw as the orchestrator and Hermes as the learning loop. But if you want one self-hosted agent that gets better at your work the more you use it, Hermes is becoming the answer.
Editor’s note: This content originally ran in the newsletter of our sister publication, The Neuron. To read more from The Neuron, sign up for its newsletter here.
The post This AI Agent Is Quietly Leaving OpenClaw in the Dust appeared first on eWEEK.
No Responses