Video: "NEW Hermes AgentOS Learn Update is ABSURD!!" by Julian Goldie on YouTube.
What /learn actually does
The /learn command is the core of this update. You describe a task in plain English — "summarise competitor blog posts and store the findings in Obsidian" — and Hermes writes a Python skill to do it. That skill is saved into your Agent OS, which means subsequent sessions start with it available without you having to describe the task again.
The more you use a skill, the more Hermes refines it. If a skill fails or produces a poor result, the next session's review loop edits the skill file — adjusting the prompt structure, tightening the output format, or improving how it handles edge cases. Skills can also be stacked: a research skill that feeds into a writing skill that feeds into a publishing skill, all chained together from a single high-level instruction.
Hermes already ships with 70-plus built-in skills. The /learn command extends that with skills shaped to your specific workflows rather than generic starting points.
Oracle: a news-watcher that runs without prompting
Oracle is a separate component that monitors sources you define — news feeds, search results, competitor sites — and refreshes its knowledge every 24 hours. The result is an agent that can answer questions about what happened in your industry today without needing you to paste in any context. Julian Goldie uses it to track SEO developments, but the same pattern works for any domain where staying current matters: recruitment, property, e-commerce pricing, regulatory changes.
Oracle stores its findings in the shared Obsidian vault, which means every other agent in your stack — including ones built with the /learn command — can read from it without extra setup.
Jarvis voice and the lead agent
The Jarvis voice assistant connects to Hermes so you can issue spoken instructions rather than typing. The voice layer runs locally rather than through a cloud speech API, which keeps your conversations private and avoids per-minute usage costs. Julian Goldie demonstrates issuing multi-step instructions by voice — "check Oracle for today's SEO news, draft a blog outline, and save it to the vault" — and Hermes executing each step in sequence.
The AI lead and email agent is the fifth new addition: a workflow that finds contact details for prospects matching a set of criteria, drafts personalised outreach emails using context from your Obsidian vault, and logs the result. Like the other components, it runs on a schedule without manual input each cycle.
What this update does not fix
The /learn command generates Python skills, which means debugging them when they fail requires enough comfort with Python to read what Hermes wrote. The self-refinement loop helps, but it does not catch every failure mode automatically — if a skill produces silently wrong output, Hermes may not detect the problem. Testing new skills against known inputs before using them in production workflows is still necessary.
Oracle's 24-hour refresh cycle is also a genuine limit for use cases that need faster updates. If you need to track breaking news or stock prices at hourly or sub-hourly intervals, a different architecture is required.
Where this connects to NordSys
Setting up Agent OS with Hermes, wiring in Oracle and Jarvis, and writing reliable skills for your specific business workflows is a configuration job that takes several days the first time through. Our AI Agents service covers the full setup — from choosing and installing the right models to building and testing the skills your business actually needs to run on a schedule without manual input.
See our AI Agents service →