Hermes 0.20: AI Agents Just Got Way Faster

Hermes Agent version 0.20 just dropped. It's called the Herald release. It changes how you use AI agents forever. You can now talk to Hermes like it's a phone call.

It answers whilst it's still thinking. You can cut it off mid-sentence just by speaking and you can say a wake word from across the room and it will start listening. And there's one feature buried in this update that I think matters more than all of that combined. I'll get to it later, but once you see it, you'll understand where all of this is heading.

Now, let me walk you through exactly what's new, how it works, and how to get it running today. First, the basics. This is from Noose Research. It's open source.

It's free, hundreds of contributors worked on the latest release and updating takes one command. So you can type Hermes update inside your terminal and you're done. If you're using the Agent OS, it's even easier. You can go to the manage section, click update, and you've got the latest version.

That's it, which means no install wizard and no settings to dig through. Now, there's a lot inside this change log to talk about. Voice, citations, speed, agent to agent stuff. And if you just heard that list, it looks like a pile of random features, but it's one idea.

So here's how I think about this whole release. I call it the open line engine. Here's what I mean. Before this update, working with an AI agent was like sending letters.

You would write something, you wait, it writes back, you read it, you write again. Every single exchange had a gap in it. The Herald release closes those gaps and now there's an open line between you and your AI agent. An open line between your agent and your apps and an open line between your agents themselves.

Three open lines. That's whole update and once you see it that way, every feature in this release clicks into place. Let's start with the first line, the one between you and Hermes. The big one is called talk to Hermes.

It's real-time voice. Hermes speaks whilst it's still generating the answer. You don't wait for it to finish thinking and then read a wall of text. So it just starts talking to you.

And here's the part that actually surprised me. You can now interrupt it mid-sentence and you just start speaking and it stops, listens and adjusts like a real phone call. So you can interrupt it mid-sentence. You don't need to type forward slash stop or stop it manually or anything like that.

You can steer it automatically using the system. So think about how you've used AI up until now. You type a prompt, you wait, you read the answer, it's wrong or it's half right. So you type another prompt or using the steer command, you wait again, that's the old way.

The new way is that you can just talk. So you can use hey Hermes as the wake word. It talks back, you cut it and cut in when it drifts. It corrects the whole back and forth that used to take maybe five type messages now happens in one conversation.

And then there is the wake word itself. So you can set your own wake word and Hermes listens for it hands free. So Hermes can sit in the background on another monitor or across the room. And when you say hey Hermes, it wakes up and starts working.

I actually tested this inside the AgentOS. So I said, hey Hermes, what time is it? And the message went straight into the chat and answered. So there's no clicking, there was no typing around, nothing.

Now I know what some of you are thinking right now. You're probably thinking this sounds technical. I'm not a technical person, or maybe this stuff isn't for me. And honestly, this update is the exact reason that belief is now backwards.

The whole point of voice is that the skill barrier is gone. If you can talk, you can run an AI agent. You don't need to even know commands anymore. You don't need to write clever prompts.

You literally speak to it the way you'd speak to a person on your team. The least technical people I know are the ones who benefit the most from this update because talking was always a thing they were already good at. Quick example, let's say for example, you're making dinner and you remember you owe a client a reply. You can say your wake word and dictate the email whilst you stir in the pan and it's drafted before you even sit down.

Now, if you want Hermes set up like this without doing any of the writing yourself, this is exactly what we built inside the agent operating system for the AI Profit Boardroom. So it's a dashboard where Hermes is already plugged in next to your claw, your open claw, and all your other AI agents. The voice agent is already wired in. You get the whole thing as a zip file, a video tutorial showing you the install step-by-step, plus a 30-day roadmap for putting Hermes to work, getting you leads and customers.

And every time Hermes ships a release, like this Herald one, we update the agent OS and you get the new version. So you're never the person figuring out an update alone, link in the comments description, orgon.com to get access. So, second part of that open line to you, voice now works everywhere. So you can send Hermes a voice note on WhatsApp and other messaging platforms and it applies with voice automatically.

It's platform aware, so it actually knows where you're messaging from and responds the right way. So your agent stops being the thing that lives in one app on one computer. It's just there on whatever app you already use all day. Let's say, for example, you're walking between meetings, you can send a voice note asking Hermes to pull together notes for your next call.

And the answer is waiting when you sit down. Next up, and this one matters a lot for anyone using AI for real work, grounded citations. There's a new citation skill and it makes Hermes backup its research with sources. Every claim is now cited.

It's kind of like how perplexity shows you where every fact comes from. So when you ask Hermes to research something, you don't just get an answer, you get the answer plus the receipts. And this hits the biggest fear people have about using AI inside their business, which is what if the AI makes stuff up? And that fear is fair.

AI models do make things up sometimes, but that's exactly why this feature exists. So when every line of research links back to a real source, you can check it. You go from trusting the AI to verifying the AI, which is a completely different thing. So let's say for example, you're comparing two supplies from your shop.

Hermes can give you the comparison and shows you where each fact came from. So you click through before you spend money. There's also a related safety layer called smart approval. So if Hermes is about to do something risky, for example, like deleting a file, it stops and asks you first.

It doesn't just plow ahead. So the agent gets more powerful. And at the same time, it gets more careful. Those two things shipping together is the right way to build stuff.

And then there's the new desktop side. So Hermes desktop now has artifacts. Everything you've created with Hermes across all your platforms shows up in one place. Images, pages, files.

And here's the useful bit. It gets sandboxed live previews. So if Hermes builds you a simple webpage or a little tool, you can preview it right there safely without opening anything on your own machine. We had Hermes generate some images through a connected tool the other day, and they all just appeared in the artifacts view ready to look through.

Now, the feature I told you about at the start, the one that I think matters more than everything else in this release, Hermes now speaks agent to agent. That's a new plugin that implements the agent to agent protocol. What that means in plain English is that Hermes can discover other AI agents, talk to them, and even be driven by them. So your agents can now work together without you sitting in the middle passing messages back and forth.

Now, let me make that real for you. In my setup, I've got multiple agents running. Claude, Hermes, OpenClaude. Before, if I wanted them to work together, I was the messenger.

I'd copy an output from one, paste it into another, explain the context again, and wait every through me. That's the old way. The new way is that I post one message and all of my agents can see it, reply to it, and build on each other's answers. I ran this the other day, posted a message, Hermes and the other agents replied, and they bounced ideas off each other, and the best ones went straight into my idea pipeline.

I didn't ferry a single message. Now, think about what that unlocks. Every agent you run has its own skills and its own tools plugged in. Maybe one is greater research.

One writes well. One handles your outreach. Hermes can orchestrate all of them. Or, and here's what's also good, is that another agent can orchestrate Hermes.

So your agents stop being separate apps and separate tabs. They become a team, and you become the person who manages the team instead of the person doing every handoff hand-by-hand. And here's why I said this is the one to watch. Voice is great today.

Speed is great today. But agent-to-agent is a feature that can bounce because every improvement to one of your agents now improves a whole group because they can all use each other. So if this protocol keeps getting better, controlling a whole fleet of AI agents gets as easy as running one. Speaking of control, there's another quality of life change that people are gonna love.

Mid-turn corrections. And here's the old way. Hermes is working on something, and you watch it go slightly wrong. Maybe it misread what you wanted.

Before, you had to stop it completely, re-explain everything from scratch, and let it start over. And that was painful. It was slow. Now you can just type the correction whilst it's working.

So you can say, do it this way instead. Skip that part. And the active turn redirects. So there's no stop command, no re-explaining.

You steer it in real time, the same way you would redirect a person halfway through a task. So you're watching it draft a post. Maybe the tone is off. You type, make it friendlier, and it adjusts it without starting over.

There's also a change to how Hermes handles long conversations. It's called improved compression. So when a chat gets really long, agents compress the older parts to make room. The problem was sometimes important stuff got squashed.

Now the compression is progress aware, and your recent conversation always survives. So the agent doesn't forget what you were just working on in the middle of the job. If you've ever had an AI lose the plot halfway through a big task, this is the fix for that. And then there's speed.

This release is fast. First token now lands in about 0.9 seconds. It used to be around 4.3 seconds. That sounds small until you feel it.

At four seconds, every exchange has an awkward pause. At under one second, it feels like the agent is just there. And this telemetry gate is 54 times faster. Plus the desktop app is quicker across the board.

So when you pair Hermes with one of the newer, faster models, tasks that used to take minutes finish in seconds. I actually gave it a guide earlier today. It learned it as a skill file, and the replies came back almost instantly. Now let me deal with the last thing I hear all the time, which is I don't have time to keep up with all these updates.

And I get it. There's a new release every week. It feels like a treadmill. But look at what this specific update actually does.

Voice means you stop typing prompts. Mid-term corrections mean you stop restarting failed tasks. 808 means you stop copying and pasting between agents. And speed means you stop waiting.

So every single feature in the Herald release gives you time back. Keeping up with this one doesn't cost you time. Skipping it does. And the update itself is just one command.

That's a whole time investment. Now here's a bigger picture. This is version 0.2. It's open source.

It's free. Hundreds of people are building on it in the open, and the release keeps landing, right? So there's new releases every single week. Last week it was 0.19, and then before that it was 0.18.

Voice, citations, agent to agent, and big speed gains, all in one single drop. Now if you compare that to where Hermes was a few versions ago to where it is today, the gap between AI agent and team member that you can actually talk to is closing fast, and it's closing in public where anyone can grab it for nothing. So here's what I'd actually do today. Four steps.

Number one, run the Hermes update. Type Hermes update or hit update in the AgentOS Manage tab. Two, set your wake word and try one hand-free request just to feel how different it is. Number three, send Hermes one voice note from your phone on an app you already use.

And four, if you've run more than one agent, turn on the A2A plugin, and let two of them talk to each other on one small task. And do those four things, and you'll understand this release better than most people covering it. I've now helped over 3,800 business owners inside the Air Profit Boarding, and plenty of them had never touched AI before they joined. A lot of them were clueless about AI agents, which is totally fair, but now a lot of them are running Hermes agent and getting the most out of it.

And this Herald release is exactly the kind of drop we build around. Inside, you get our Hermes agent operating system as a zip file where Hermes sits alongside your other AI agents with voice, memory, workflows already wired in. You get the 30-day roadmap for turning Hermes into something that saves you hours and brings in leads and customers. You get daily step-by-step tutorials, and the new ones will walk you through the Herald release features one by one.

So you can talk to Hermes, wake words, citations, agent to agent, and there are four coaching calls every single week where you can share your screen. You can ask about your own Hermes setup live and get unstuck on the spot. There's also a prompt library, a member map, so you can find other people near you running Hermes in the AgentOS. And there's always someone online 24-7, so you can get help at any hour inside the community.

Link in the comments description, or go to theaiprofitballroom.com to get access. That's the Herald release. Your agent talks now. It listens for you across the room.

It cites its sources. It asks before doing anything risky. It's faster than it's ever been, and it can finally work with your other AI agents instead of next to them. The open line is here.

The people who pick up first are gonna be miles ahead. Thanks for watching.