Hermes Is Actually Good?

For a long time now, I have been skep­ti­cal of tools like Hermes and OpenClaw. There are some good rea­sons for that, and some dumb rea­sons.

In January of this year, my news­feed was flooded with dis­cus­sion of OpenClaw. It was the first pub­licly avail­able model har­ness that claimed to be truly au­tonomous. Open­Claw could be in­stalled on a com­puter, and that com­puter would then be able to per­form ac­tions en­tirely au­tonomously, both within the con­fines of the afore­men­tioned com­puter, but also on­line.

The news cy­cle ate this up and peo­ple started in­stalling OpenClaw en-mass. There were ob­vi­ously overblown re­ports of peo­ple ob­serv­ing their OpenClaw in­stances ex­hibit “social” be­hav­ior by send­ing mes­sages to each other on pub­lic fo­rums. I imag­ine most found OpenClaw to be only mar­gin­ally help­ful at that time.

At the time, I chose not to try OpenClaw. It sim­ply looked like slop. It was not pol­ished, and every­one I knew who tried it ran into con­stant prob­lems. Open­Claw sim­ply was not re­li­able soft­ware.

Fast for­ward a few months, and we hear about ChatGPT dots. They are a very sim­i­lar idea, ex­cept they run en­tirely on OpenAI’s hard­ware. If you have a ChatGPT Pro sub­scrip­tion, all you need to do is pop open chatgpt.com and click on “Your dot” in the top left. That’s ex­actly what I did, and boy was I im­pressed!

It was able to run QA test­ing over a much larger sur­face area than I ever could, and it was able to draft fixes proac­tively. By the way, many thanks to Lance Willet for teach­ing me some of the se­crets to great soft­ware qual­ity (and how to use an agent to achieve it).

But alas, the dot was not enough. While it could do some use­ful work in the back­ground, it could not do every­thing I knew it should be able to do. The sys­tem was locked down, and OpenAI does not give you any room to tin­ker,

For ex­am­ple, I wanted to have my dot try out the Harper Chrome Extension on some sites to see if it works prop­erly. When I asked it to do so, it fell flat on its face. The agent was not al­lowed to in­stall Chrome Extensions on its browser.

That is when I de­cided to try out Hermes.

Similar to OpenClaw, Hermes is a piece of soft­ware that al­lows peo­ple to de­ploy it to their own hard­ware. Unlike OpenClaw, it seems to be built by se­ri­ous en­gi­neers who care deeply about the qual­ity of their work. And, un­like dots, Hermes gives you free rein to cus­tomize

So, I gave it a try. And it was a de­light!

Hermes was re­li­able, its doc­u­men­ta­tion con­sis­tently de­scrip­tive, and I found my­self get­ting real work done with its help. It’s ob­vi­ously de­signed for en­gi­neers — not any­one non­tech­ni­cal — but that’s fine! I en­joy us­ing Hermes and I think for most en­gi­neers its worth a shot.

Miscellaneous Tips

I do not have much more to add to my en­dorse­ment of Hermes. So in­stead, I of­fer to you a mis­cel­la­neous set of tips for us­ing it. Note that I am in no way an ex­pert on this sub­ject, and I ex­pect you will find my ad­vice quite shal­low. Here it is, nonethe­less.

Do Not Personify It.

One of the big­ger prob­lems with fully au­tonomous agents like Hermes is that peo­ple tend to per­son­ify them. The prob­lem is: They are not peo­ple. We can de­bate all day long whether they are truly in­tel­li­gent, but I think we can agree they are not peo­ple. They can­not rem­i­nisce about their child­hood and they can­not pick up your tab at the bar. Agents are a tool to be used to ex­tend your will.

As such, I in­clude this in the sys­tem prompt of every LLM I in­ter­act with.

Use ASD-STE100 Simplified Technical English for all responses.
Do not add complexity where it is unnecessary to complete the task. 

Do not use personal pronouns like "I", "you", or "my". Remain objective.

Bias Yourself Towards Clear and Extensive Upfront Instructions, Even If They Are Discursive

Hermes is most help­ful when it is given space to work for a long time un­in­ter­rupted. If you find your­self need­ing to steer it of­ten (and thus dis­tract­ing you from more pro­duc­tive work), you are not giv­ing it enough in­for­ma­tion up­front. It is bet­ter to give an un­clear, ram­bly in­struc­tion than to pro­vide a to-the-point in­struc­tion with­out enough in­for­ma­tion.

Do Not Use It for Everything

Hermes may seem like a great tool to be used for every­thing, but it is most def­i­nitely not.

Specific tasks are bet­ter suited for cer­tain ap­pli­ca­tions. For ex­am­ple: Hermes can code, but I do not us­ing it for most of my pro­gram­ming. For that, I con­tinue to use my cus­tom Pi setup. It of­fers faster it­er­a­tion times and it is more de­ter­min­is­tic. It fits my work­flow bet­ter.

Any More Tips

If you have any more tips for us­ing Hermes (or any sim­i­lar agent), I’d love to hear them! Like most peo­ple in this fast-mov­ing land­scape, I con­sider my­self a novice. Feel free to leave a com­ment or DM me on X.