How I work with Hermes

I use Hermes for much more than asking questions. It's where I work through ideas, research something I don't understand yet, and coordinate work with other agents. Auren is my personal assistant in that setup. Hermes itself is built by Nous Research.

The part I find exciting is being able to start with a problem instead of a perfectly prepared task. I can say what I want to improve, look at what already exists with an agent, and work toward a useful next step. That next step might be code, a document, a design or simply a better question.

Find out what we already know

Before building, I want to know whether we've worked on the problem before and what the current state actually is. For code, that means reading the repository and checking its existing tests and conventions. For research, it can mean searching my wiki and then opening the original paper or documentation.

The wiki is useful because it connects ideas I've already spent time on. A paper about memory might relate to a tool I'm using, an earlier comparison or a question I still haven't answered. That gives us a better starting point than another search from scratch.

It doesn't make the answer automatically current or correct. I still want the agent to check the source when the detail matters.

Turn the idea into work

Once the goal is clear, I work with the agent to decide what should happen and how we will know it worked. If the task can be split, different workers can take different parts.

For a website change, one agent might check the existing pages and another might work on the copy. A coding agent could handle a focused implementation task in a separate workspace. The useful split depends on the work, not on how many agents I can launch.

I use tools such as Cursor and Codex for coding alongside Hermes. I want the freedom to choose a model or worker for the job and change that choice when something else works better.

Check what came back

The handoff back matters as much as the handoff out. A summary helps me understand what the worker attempted, but I want to inspect the actual result. For code, that means the change and its tests. For writing, it means reading the copy and checking the claims against the sources.

I also use a separate reviewer that didn't produce the work. That gives the result another check against the original request. When something is wrong, the finding should be specific enough to fix.

For consequential actions, such as publishing or moving money, the workflow also needs my approval. Doing useful preparation and having permission to act are different things.

Keep what we learned

When we work out a procedure worth reusing, I want it saved as a skill the next agent can load when it needs it. A project decision stays with the project. A research finding belongs with its sources. That way the next session can benefit without carrying the whole history in its prompt.

I keep refining this because it lets me take on work I would otherwise have to do one piece at a time. There's still judgment involved in choosing the task and checking the result. That's work I want to spend more time on.