Your second brain

If you are trying to use AI agents to manage workflow you need understand what works for you. Pic/iStock

Personal AI agents are rapidly becoming the next big category in consumer tech. Not chatbots that answer trivia questions or generate anime profile pictures. Actual software agents that can operate your computer, browse websites, automate tasks, message people, and organise your digital mess while you pretend to be productive. The dream is simple. You tell the machine what you want, and it quietly handles the boring stuff in the background. The reality, however, has been slightly chaotic. Most AI agents today feel like interns with administrator privileges. They can be brilliant one moment and catastrophically stupid the next.

The rise and fall of Openclaw

That is what made OpenClaw interesting when it first appeared. OpenClaw was one of the first open-source agents to genuinely feel useful outside controlled demos. It could navigate apps, control browsers, manipulate files, and automate repetitive workflows using natural language prompts. Suddenly, you did not need to build complicated automation chains or learn scripting just to make your PC do useful things.

You could simply tell it what to do. "Organise my downloads folder." "Summarise unread emails." "Pull invoices from Gmail and save them to Dropbox." And surprisingly often, it worked.

The problem was that OpenClaw tried very hard to remember everything. Every interaction, every context window, every workflow, and behavioral pattern slowly piled up into long-term memory. Initially, this felt impressive. After a while, the system became slower and more cumbersome. Memory retrieval got messy. Old context polluted new tasks.

Enter Hermes

That is where Hermes takes a different approach. Hermes still uses persistent memory, but it treats memory more like a curated notebook than a storage locker. Instead of endlessly preserving context, it aggressively summarises, prioritises, and trims information it considers less useful. The goal is not to remember everything. The goal is to remember what matters. That sounds like a tiny distinction until you actually use it.

For example, if you repeatedly ask Hermes to open Slack, check your calendar, and summarise missed project updates every morning, Hermes can slowly turn that into a reusable workflow. But unlike OpenClaw, it is less likely to drag irrelevant historical baggage into the task six weeks later. In other words, Hermes tries to age gracefully.

There is also a noticeable difference in how the two systems approach autonomy. OpenClaw often felt heavily prompt-driven. Hermes appears more willing to break larger goals into smaller actions on its own. Sometimes it feels genuinely clever. Other times it feels like watching a very confident person assemble
IKEA furniture without instructions. Still, the direction is interesting.

One particularly useful feature is the ability to write self-written skills. Hermes can observe repeated actions and automatically create reusable task modules. Think of it as the agent slowly building shortcuts for itself instead of forcing the user to manually configure everything. That makes Hermes feel less like a chatbot and more like a junior digital operator learning routines over time. Of course, giving an AI agent direct access to your desktop is also slightly insane. Thankfully, Hermes seems to understand this.

The platform separates sensitive operations into permission layers and sandboxes many risky actions. Shell commands, credentials, and external messaging usually require approval depending on your setup. There are also activity logs so you can inspect what the agent actually did after the fact. That said, no amount of security design can protect users from poor judgment.

Do not give experimental AI software unrestricted root access and act shocked when it deletes your movie collection while trying to "optimise storage". Hermes can automate tasks. It cannot automate common sense.

There is one small catch. Hermes is primarily designed for Linux environments. Windows users will need WSL2, Microsoft's Linux compatibility layer, before installation even begins. Thankfully, WSL2 is much less scary than it sounds and can be enabled with only a few commands.

A screen shot of Hermes workflow

If you are on Windows
(Linux and Mac OS can skip this step), start with:
wsl --install
Then (inside WSL, Linux and Mac users start directly from here): curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
Finally: hermes

Follow the onscreen instructions, and you are done

If you are even remotely nervous about running an AI agent directly on your main operating system, use a virtual machine first. Seriously. Breaking a VM is educational. Breaking your actual work machine is character development. That said running any agent like this is risky so if you are uncomfortable running this on your own PC, it can also run on a cheaper machine if you are willing to take the AI stuff online to maybe DeepSeek or OpenAI etc, Hermes provides on screen instructions during the initial setup.

There is also a migration shortcut called hermes-claw-migrate for OpenClaw users who want to migrate existing workflows to Hermes.

Hermes works similar to OpenClaw in many ways but the architectural difference means that Hermes still works great on day 90 while OpenClaw fails.

Persistent AI Agents honestly feel like magic and it is very impressive how these systems manage to learn and evolve. Something worth investing a bit of time in learning how to setup, even if it is only for educational purposes.