Hermes Agent - The good, the bad, and the ugly

I spent last Friday jumping on the latest AI Agent bandwagon.

Allow me to introduce you to the latest hyped up tech in an already crowded world of AI agent software: Hermes Agent.

It's touted as a self-learning agent. I think a more accurate term is self-documenting.

I did have some serious issues with it.

Even if you turn on manual approval for commands, it still only asks for approval for a select few commands, which led to it running a wide variety of Python scripts the model coded on the fly that I had no idea what they did.

Obviously, this is a massive security flaw, and before you ask, yes, I was running this in a container to prevent it from blowing up my desktop.

It also seemed to rewrite skills at random; skills I had not yet asked it to use in any way.

Randomly altering documentation for tasks it had not been instructed to work on isn’t what I would consider “Self-improving”.

I was able to get it to connect to Unity using the Unity MCP server to see if I could get it to do anything at all with Tech Debt, like perhaps QA it so I can get the next round of playtest out of the door.

Sadly, it wasn’t able to do much. I asked it to make a copy of a button… It failed pretty hard, so I am not optimistic for significant contributions any time soon.

I wish I could tell you what the AI hype bros are screaming from the rooftop, that “This new software will change your life”, but I cannot.

What I can tell you for certain is that if you play with any of these AI tools, DO NOT GIVE THEM ACCESS TO YOUR COMPUTER; keep them in a container or on a remote virtual server with safeguards.

While there is a small chance these things could give you some productivity boosts, there is an infinitely larger possibility that they will brick your OS or publish your sensitive documents to a publicly accessible website for malicious parties to abuse.

The agent actually did push text files that could have been sensitive to a public site only a few hours after I started working with it, but again, since I was in a container, it was nothing sensitive.

With all that said, I am going to play with it a bit more and see what I can get it to do. I’ll let you know how it goes.

What AI productivity tools are you using?