Video: "I Built Jarvis With Hermes Agent, Here's How!" by Julian Goldie on YouTube.
What the build involved
The stack is straightforward: Hermes Agent as the operating layer, a capable model with a large enough context window to hold extended conversations, and Hermes's built-in voice mode for input and output. Julian's setup uses the TTS layer for spoken replies and microphone access for input, both configurable in the settings panel. No third-party voice software is needed.
The session moves from install through model connection, system prompt configuration and skill loading — everything in a single sitting.
The system prompt — where the personality lives
The most important step is the system prompt. This is the standing instruction that defines what the agent is, what it knows about you, and how it should handle requests. In Julian's build, this is where the "Jarvis" behaviour comes from — not any special code, just a well-written prompt that gives the agent a clear role, a working style, and a set of priorities.
This is the step that separates a generic Hermes install from a genuinely useful personal assistant.
What it looked like in use
The demo shows the assistant holding context across a working conversation — referring back to earlier tasks, taking on instructions rather than just responding, and speaking its outputs. Julian runs it through several real task types: requesting a search, getting a spoken summary, dictating a brief and having the agent structure it. It does not feel like talking to a chatbot.
That said, the agent needs clear instructions. Ambiguous requests produce the same uneven results they would over text.
What is still worth planning for
Three things to be aware of before building your own. First, voice latency: there is always a gap between speaking and response, and it varies with your hardware and connection. Second, memory resets: Hermes does not automatically retain memory across restarts unless you configure a persistent memory layer — worth setting up before you rely on it.
Third, the system prompt needs iteration. The first version rarely captures everything useful. Expect to refine it over the first week of use.
The broader implication
The more useful signal from this video is not that Julian built something cool — it's that the complexity of building a working personal AI agent is now at a level a non-developer can manage in an afternoon. The barrier has shifted from technical to conceptual: you need a clear idea of what you want the agent to do, and a bit of patience with the prompt. That's a genuinely different situation from two years ago.
Where this connects to NordSys
We help businesses work out what they actually need from an AI agent — then build and configure it properly, with persistent memory, the right skills, and a brief that holds up in daily use. If you want a Hermes setup that does useful work from day one rather than a science project, that is what we do.
See our AI Agents service →