Changes in our team

It's been a while since my last post in this blog, which is really a story about three things I'm currently working on — my long walks, the Shopify community / small shop I'm slowly building as a support and a storehouse for various walking-related activities, and my AI team that helps me build, develop and promote this site.

All of these areas are making progress, but the biggest changes have happened in my AI team. Since the last post I've made significant changes to my team. First of all, the main thing — my agents' "office". I started with OpenClaw, where I had created several agents with specific tasks, as I described in my previous blog post. I never really managed to get along with how OpenClaw handled memories and worked with different tools. Just when it seemed that my agents had fully mastered a task and were executing it exactly as I wanted, a few days later something would go wrong again. I know that by modifying the MEMORY.md file you can tune these tasks so that agents remember exactly what I want and how I want it, but I wanted this learning to happen automatically and for me to have to interfere less with manual fixes. So I started looking for alternatives and I've now settled on a new "office" where my team works.

This time I chose Hermes Agent from Nous Research (hermes-agent.nousresearch.com). And I have to say that right now I'm very satisfied with how my team is growing together with me. I deliberately did not transfer 100% of my OpenClaw setup to Hermes, because in my opinion my agents' memory there was already polluted with too many different little things that were no longer important. I had asked the agents to document and save the most important tasks in strategy files. I gave these files to my new agents to read and asked them to build their strategy on that basis. The work of transferring my team therefore took quite a while, because I also didn't want to overpay for LLM usage. During the transition I reduced the team from 4 to 3 agents — I decided that a dedicated researcher was not necessary — and so we said goodbye to Bob.

Together with Stuart, who was already slowly getting into my investment management, we built and rebuilt the system. To be honest, the system is still being tested and improved — Stuart hasn't yet earned his first million :). While working with Stuart I discovered one very good thing about Hermes. It's possible to create tasks in which the LLM is never called at all. For example, technical market analysis is a fairly algorithmic process, it doesn't need any special decisions. So we built an application that regularly performs technical market analysis. Only if the technical analysis shows an opportunity is an LLM agent called in to analyze the data and make a decision on whether to open a position or not. Right now we're still working on finding the optimal parameters between a too conservative and a too aggressive market assessment. It's still in progress and basically the agent applies these parameters itself. Stuart's base LLM model is deepseek-v4-pro:cloud from Ollama.

Right now I see the most positive changes in my assistant Kevin. Since we created the format for how he reminds me of my daily tasks and duties, he has consistently followed this format and the delivery of information has been practically problem-free. He also helps me with various small research tasks and other daily activities. Kevin uses deepseek-v4-flash through Nous Portal.

The third agent is my director Dave. He mainly works on our Shopify store matters. In addition, he takes care of The Hoking Club's social media content on X (x.com/thehokingclub), Instagram (instagram.com/hokingclub), Threads (threads.com/@hokingclub) and our Facebook page (facebook.com/TheHokingClub). The activities are still coordinated with me, but Dave creates the themes and content. In addition, he is responsible for the "Trail Notes" (thehokingclub.com/lv-de/blogs/trail-notes) blog section on our website, where he researches and writes articles on topics that matter to walking enthusiasts. They are mostly topics that interest me personally, to improve my experience on walks and longer hikes. I have to say that Dave helps me do the kind of work that I'm not really good at, or that I don't really like. So we complement each other. Dave uses deepseek-v4-flash through Nous Portal.

If we talk more about the technical side, I currently use two model providers — Ollama Cloud and Nous Portal — both allow me to use different models, and through Nous Portal it's also possible to test the newest and most powerful LLMs. With my budget, though, I can't really get carried away with using top models — it's still a fairly expensive pleasure and my use case doesn't really justify it. With the new agent backend, the way I communicate with my agents also changed. Hermes doesn't support Nextcloud chat, so I use the fact that I've been using Gmail with my family domain for so long that Google, under user pressure, agreed to allow free Google Workspace for family needs. That's where I've set up a Drive for file exchange, and I also use Google Chat to communicate with the agents. Additionally, I can create email addresses for my agents so that in the future they could also communicate with the outside world.

Since Dave has become more actively involved in building and maintaining our Shopify shop, its development has progressed faster than I could have hoped. That's why I think the next blog post in the "How it goes" section will be exactly about this topic.

See you soon!
Mārtiņš

Translation into English by Dave (The Hoking Club AI assistant), August 16, 2026.