01What the Hermes voice agent is and how it works
The Hermes voice agent is called Hermes Apollo. It lives inside Julian Goldie's Agent OS dashboard, where it talks, listens, sees your screen and remembers what you have done with your other AI agents.
In the demo, Julian says open Google and it opens Google. He asks for SEO keywords and gets a list right away. He asks it to open his Obsidian notes and it reads his latest projects back out loud.
When he asks if it can see his screen, it names the page he has open.
This episode shows the whole thing working:
- The voice agent.
- A background mode that gives a daily briefing.
- A pixel office where you watch AI agents work.
- A video agent that turns one idea into a finished video.
02Streaming voice and barge-in
Voice mode used to work like a voicemail. You speak, wait for the whole reply to be written, then listen to one long audio file. That felt slow.
Hermes added streaming voice. You get a real conversation, wake words that run on your own device, and a standard way to talk to other AI agents.
Here is how the streaming works. Hermes gathers words from the AI until it has a full sentence of at least 20 characters, then says it out loud. So the reply starts talking while it is still thinking.
You can also cut it off mid-sentence. This is called barge-in. In the demo, Julian said stop in the middle of a long list and it stopped.
03How Hermes Apollo is set up in Agent OS
Hermes Apollo sits inside the Agent OS dashboard. It does not use a separate paid API key. Julian already has a ChatGPT plan, so he hooked up the ChatGPT voice command line tool and plugged it into Agent OS.
One plan he already uses now lets his whole dashboard talk.
You can change the voice whenever you want. Options include a deep voice, a calm voice or a butler voice.
You can also shrink it down. It sits in the corner ready to go, and clicking it pops it back open. Or you can switch to the full Hermes Apollo tab for the big view.
If you install a tool called faster whisper, the speech to text part works with zero API keys.
04The big view: chat history and things you built
In the big view, you get the full chat history. Every voice chat you have had is saved, so you can scroll back and see what you asked last week.
Below that, you can see everything you built with it. Julian asked it to build a simple to do list app, and there it was. You can open it full size right from that page.
So you talk, it builds and it keeps a record.
There is also multitasking. Every tab in Agent OS has a small pop-up on the right side. While text Hermes is busy writing a long reply, such as a 30 day content plan, you can open Hermes Apollo and ask voice Hermes to find five keyword ideas about AI agents for small business.
Two jobs run at once, so you are not staring at a loading bar.
05Screen sharing and wall mode
You can share any page inside Agent OS with Hermes Apollo. It reads what is on the screen and understands what you are working on.
In the demo, Julian had the Hermes chat page open with a chat about SEO keywords. He asked if it could see his screen, and Apollo said yes, naming Agent OS, the Hermes page and the chat about keywords and content ideas.
This matters because you stop having to explain things. With a draft post open you can say look at this and tell me what is weak.
Julian calls the background mode wall mode. The voice agent sits in the background, on a second screen or on the wall, and still works in real time. You can ask for a daily briefing:
- What your agents did overnight.
- What is on your list.
- What is waiting for you.
You could also ask what questions people posted in the community overnight and hear them read out while you make coffee.
06Memory with Obsidian and shared agents
Memory is the part that makes the Hermes voice agent strong. Every time you use Hermes Apollo, it writes into the Agent OS memory system. That memory lives in Obsidian, a free notes app that keeps your notes as simple files on your computer.
Every other agent inside Agent OS reads from the same notes. If you tell the voice agent this week you are focusing on welcome posts for new members, your text, SEO and content agents all know it too.
It works the other way as well. Julian has a daily updates automation that runs on its own and writes what it did into memory. The voice agent can read that.
If you already have notes in Obsidian, connect them. A voice agent that knows your projects is far more useful than one that does not.
07Starnet: the pixel office for your AI agents
Starnet is a pixel office for your AI agents. Picture a little old school video game office with tiny pixel people walking around, sitting at desks and on a sofa. Each one is a real working agent.
Click on the marketer and you see what it is working on, its history and every run it has done. Click on Julian's AI agent and you see the same. You can give any of them an order or add a new agent.
You can even have more than one office:
- An engineering office where things get built and tested.
- The main office where Julian's AI agent usually works.
- A corner where the web designer agent waits on the sofa for a job.
Why a game looking office? It makes your agents easy to check. You do not read logs.
You just look at who is busy, who is free and what they finished.
08Limits, first steps and who it is for
Julian spends about three to four hours a day improving his Agent OS, so building it from scratch would take a long time. A ready made setup is the shortcut.
Voice agents can mishear you, and screen reading can miss things. That is why memory and history matter. You can always see what it did and check it, so you stay the boss.
Here is how to start:
- If you are new to agents, start with one small voice job, such as read me my notes from yesterday.
- If you already use Hermes, turn on voice mode.
- Try barge-in by interrupting it.
- Connect your Obsidian notes.
- If you run a team, ask whether you can see what your agents are doing at a glance.
Julian built this one piece at a time: voice one week, memory another week, the pixel office after that.
