How My Stack Shifted Again - From Hermes Agent to a Harness of My Own

I moved my stack again.

In May it was VS Code with the OpenChamber plugin, and me steering every line. In August that gave way to Hermes Agent, and I wrote about how I stopped opening the editor. Now the ground has shifted a second time.

The short version: I am working inside the DeepSeek Harness for projects and development, and then using Marimo Desktop to bridge the gaps.

What Changed

OpenCode and the third-party harnesses are behind me now. For a long stretch my setup was a pile of parts that each did one job well and knew nothing about the others. I spent more time and money than I want to admit patching the seams between them.

Hermes Agent was the first setup that felt like one piece. It held my context, it remembered how I work, and it kept the projects in one place. That is why I wrote about it a few weeks back in How My Coding Workflow Changed - From VS Code to a Team of Agents.

Marimo Desktop is the next step in the same direction. It is my own agentic harness, so the fit is closer. It’s not built to handle my coding. That’s what the DeepSeek Harness is for.

It was built to bridge the gaps and automate tasks while ultimately being my assistant. Right now, it has tools, skills, memory, identity, and agent profiles.

My first agent is Hoshi, my blogging and social media assistant. The agent learned how I write by analyzing my previous posts and building out a voice profile and reusable skill. Now I can turn to Hoshi when I need content.

I always proofread, edit, and compose the final draft, so it’s always me behind the wheel. I know what goes in, and what comes out, while keeping the base organized, simple, and customized to me.

When something about the workflow annoys me, I change it, because I am the one holding the wrench.

The Two Model Setup

Underneath it all sits DeepSeek. I run it direct through the API with DeepSeek v4.1 Flash, and the cost is super low. That matters more than it sounds. A low cost model I am happy to hand work to all day gives me peace of mind that I can run long tasks, dive into bugs, and implement new features without worry.

During peak times I switch over to OpenRouter and run the older DeepSeek v4 Flash 0731. Same family, two roads to it. The older model is one I already know, and there is comfort in having it on the shelf when I want it. Especially for smaller tasks that do not take as much reasoning effort.

Not only does this save money, it also saves time. I just make sure to keep my longer more complicated tasks for DeepSeek v4.1 Flash.

More Agentic, Both Ways

DeepSeek v4.1 Flash feels more agentic than what I was on before. It gets more done without me stepping in to steer it.

That is a good thing and a bad thing, and I mean both halves.

When a model needs no steering, it also does not stop to ask. On some work that is exactly right. A refactor with clear edges and a well-defined prompt. A batch of small fixes where I already know the shape of the answer. Other work wants a hand on the wheel, and there I would rather do the steering myself.

So far it has been really good. The trick is choosing when and what to use it for. That is true of every tool in the house, and this one just makes the choice obvious.

Why It Feels Good

Two things stand out. The cost, and the quiet.

The cost is easy to explain. Low through the direct API, and a fallback for the busy hours. I can work the way I want without watching a meter.

The quiet is harder to put down. For a long time my setup was a pile I kept re-gluing. Now it is one agent harness I built, one model I trust, and one modular coding harness. Fewer moving pieces means less time tending the machine and more time on the actual work.

It has been a breath of fresh air, and it feels good.

If you are mid-shuffle yourself, I know the feeling. Try starting from a minimal point and then building up from there. Watch what you stop worrying about and what really matters to you. You don’t need 1,000 plugins or a million lines of agent code to get what you need done. You just need something that works. For you.

Anthropic’s Building effective agents is a good starting place to learn about the importance of starting small. Building up with what you need, then taking flight.

For now I am staying here a while… I think :D

The landscape is moving fast, we will have to wait and see what happens next month.

Until then, don’t stop building.

---

P.S. Marimo Desktop is my own agentic harness, early days and currently under development. DeepSeek v4.1 Flash through the direct API is my daily driver. Peak hours go to the older DeepSeek v4 Flash 0731 through OpenRouter.