Hermes Agent

Hello again my friend,

This is a ‘Product Building’ post, where I cover the best products in the industry, and it’s one I’ve been looking forward to now for a while. Before sharing the details, I want to start with a story.

A few weeks ago I woke up, made my coffee, and opened my laptop to about 40 finished tasks. Bug fixes, UI changes, little cleanups on an app I've been building to manage my expenses and taxes. I didn't do any of them. I was asleep.

Before bed the night before, I had Sol, one of the AI models from OpenAI (ChatGPT), make a plan and split the work into tasks on a board. Then, I went to bed. Sol handed the work to a different (cheaper) model, that worked for the entire night, non-stop.

Total bill for the night: around two and a half dollars.

Some of the work needed fixing the next morning, sure, it wasn’t perfect. But I went to bed with a to-do list and woke up with a review list, and that is a very different way to live. Especially for me, having been very stressed and struggling to manage the endless stream of tasks.

The software is called Hermes Agent. If you check it out, I use their desktop app specifically.

It's rare for me to talk about a product like, this. The last was beehiiv, which I used to start my last business, before joining their team full time, and helping scale from 10 people to 130. beehiiv was a major unlock for me and my career.

So when I say this is the first software since then that felt like that big of an unlock for me, I don't say it lightly.

Let’s lock in.

Compare that lovely morning to where I was earlier this year, you may recall I shared some of that in a previous post (the trap in What a time to be alive).

I was paying for the best models, and I used them constantly. But still, I was drowning, and nothing ever felt finished. I’d be repeating the same instructions to a chatbot that forgot me every morning. Half a session would be spent dicking around with setup, with tools, with connecting this thing to that thing, instead of doing the actual work.

And always, always sitting next to the AI, telling it what to do next, because the moment I walked away, everything stopped.

The models kept getting better, and my weeks didn't. More output, more drafts, more motion.

This post is about that gap, from constant stress, to a feeling of real progress. Yes, I’ll talk about the product, but this is the outcome I hope you find from it, if you’re feeling the way I did.

Hermes agent is a harness. That word means nothing yet, so here's an example.

In Iron Man, Robert Downey Jr. plays Tony Stark, a genius inventor who builds the iconic Iron Man suit. To power the suit, he creates the Arc Reactor, a power core that gives the suit the energy it needs. That core evolves over time, and whole scenes in those movies are dedicated to upgrading it.

The power core is like an AI model. When someone talks about Claude's Fable or Opus models, or ChatGPT's Sol model, they're talking about cores.

The suit itself is the harness. Tony builds different suits for different purposes. Some fly faster. Some are massive and strong (one fights The Hulk). Some are loaded with weapons. The core provides the power, but the suit dictates what that power can actually do. (Forgive me if you're a die-hard Iron Man fan. You know what I'm saying).

A power core without a great suit (again, harness), isn’t Iron Man. It’s Tony Stark with a lot of energy, but nothing to point it at. It doesn't fly, it doesn't lift, it doesn't protect anyone.

That was me earlier this year. I had the core. I was the guy holding a glowing circle, busy and impressed and not actually flying anywhere or doing anything that cool. Having it was an upgrade, of course, but still lacking in a way.

I won't claim that's a knock on ChatGPT or Claude. They’re great.

But historically, the Fable model is amazing, but the Claude Code harness hasn’t been (at least for me).

Hermes agent is a full suit, and you get to choose which core to put in it. So yes, I can use Fable, or Sol, or DeepSeek v4 Flash (another, cheaper AI model), but keep the same harness. And there’s a reason I want to keep Hermes.

Every so often (around 20-30 chats back and forth), Hermes internalizes something about working with you. Like something you’ve told it, instructions, critique, etc.

Early on, I kept telling it to stop agreeing with everything I said and to push back more. At the time, all the models did that. After a few hundred turns (this came up ~10 more times), eventually, it was just more critical, and I didn’t have to remind it.

After enough turns, your instance morphs into something that fits you. Your style of working, your preferences, your coaching. It compounds quietly, the way a good working relationship with a person does.

You don't notice it happening day to day. You notice it a month later, when you start a new conversation, and it’s as if there was no gap with the last one. No re-training, no 3 pages of prompting to bring it up to speed, no asking the last session to write a recap for the new session. It just knows.

Remember the drowning version of me, repeating the same instructions every morning? That's the first thing that got fixed.

And, all of those memories/knowledge stay, even if I switch from Claude to ChatGPT, to something else. Normally, if I go from Claude Code to Codex, I have to re-teach everything. It’s like a brand new employee.

Hermes has a model picker, and it's exactly what it sounds like. At any time, even mid-conversation, you can go from Claude's Fable to ChatGPT's Sol to something far cheaper like DeepSeek (whatever fits the task/budget).

But, everything it’s learned about you stays. If you switch from Claude to ChatGPT inside Hermes, you don't re-teach anything, and you don't play messenger between two apps that have never met. It's all built in.

This is also why I think cheap models are about to get very interesting. In my observation, a heavy task on a frontier model like Fable 5 runs somewhere around $3. The model that worked my overnight shift cost about $0.04 per task.

That's roughly 75x cheaper, and because the harness carries the memory and the skills, the cheap model arrives already knowing how you work. The core is smaller, but it's dropped into a suit that's been fitted to you for months.

And to be clear, part of the reason why it was so cheap is because Nous (the company that built Hermes), had negotiated a 90% discount on the model at the time. So it’s not always like that, but they do it fairly often.

I wrote before about what happens when companies don't manage their AI spend. This is the other side of that coin. The tools for spending dramatically less are showing up, and most people haven't noticed yet.

For a long time, there was a lot of talk online about a “team of agents,” and I never really understood how people actually did it. This is the solution I landed on.

A profile is a completely separate identity with its own purpose, memories, skills, and connections. A 'coder' profile, a 'content' profile, an 'accountant' profile, is effectively different, specialized employees. But, you also get a default, so there’s no major setup upfront, this is only if you want it.

Each profile can get a different model, and assigned different tasks. These profiles can talk to each other (Hermes has Agent to Agent conversations), so they can work without you. If you prefer, you can add multiple profiles into a group chat.

Hermes has bot mode (similar to Grok bot), if you’ve never heard of it. You can create a new group, add your profiles, and talk to them. You can do it on the desktop app they provide, or in Discord.

Now you might be wondering, are profiles the reason I could get a full night’s work done remotely, while the computer was off?

No. That’s part of Nous’ Cloud offer. Essentially, you can buy a digital computer from them, so when your laptop is off or closed, they’re all still active and working. It’s about $1 a day (depends on the one you get), and if you’re not using it for a month, you can power it down, and they’ll just charge pennies a day.

I added profiles to the cloud computer I bought from them. And so whether it’s the desktop app, or Discord, they’re never off. Like the other day I was at an event, and I got a message on Discord from Hermes asking me if I could approve the plan for that evening before handing off a few dozen tasks for another profile to work on. I read it and approved it via Discord.

That’s how you get a team of agents working, and keep them on overnight even if your computer is off.

On day one, it feels about the same actually, it’s just a different feature set (skills/mcps are pre-loaded, self-learning loop, model picker, etc). The real difference appears over months.

The difference shows up over months, for all the reasons I described above.

Think of Grok bot, or Claude, or ChatGPT like Apple. Everything’s in the box, it’s easy, there’s little or no customization required (or possible). It’s there, it’s done for you, go ahead.

Hermes is more about the flexibility and the options. The entire platform is open source, meaning you could host it yourself without paying them anything, and just pay for the models you use. It’s so flexible, you can do more or less anything with it.

And their team ships at a pace that honestly reminds me of us at beehiiv. In the months I've used it, they've put out over a dozen large releases and thousands of smaller fixes, including ones I've flagged to them personally.

But, I find that their desktop app makes it feel just as easy, but most technical people don’t use it, so it’s not talked about as much.

Anyway, I can't tell you Hermes is better than ChatGPT or Claude for you. That's too subjective for a blanket statement.

They are not paying me for an ad. I'm writing this because you, my faithful reader, come to me for thoughts on how to turn hard work into light work. This is a piece of software that has genuinely helped me do that, and I suspect it will keep helping more over time.

You can download their desktop app here.

This is the always on computer I bought from them if you want to try having AI work for you while you sleep.

Thanks, as always, for reading.

Darwin