• 2 min read

Playing Around with Agentic AI

How I turned a rarely used gaming laptop into an AI agentic infrastructure.

The AI agentic platform is widely talked about, and that sparked my interest in the possibilities it could give. My laptop has been lying around without much use, so I had to make it useful. Considering the extent of possibilities and making use of my available tools, I decided to turn it into my local AI agentic infrastructure.

The tools/apps I'm exploring are:

  • Llama.cpp: To self-host an LLM model.
  • OpenClaw: Personal Agentic AI
  • Hermes Agent: Personal Agentic AI
  • Firecrawl: Self-hosted browser capabilities.
  • Honcho Memory: Memory layer for agentic AI.
  • qmd: Search engine for my local docs.
  • Obsidian: Managing my local docs/notes.

Starting from the OS, I use Arch as the base OS, and a QEMU VM running Debian 13 for the AI system. It is crucial to compartmentalize my device for security purposes, so I went that path. Given the enormous experimentation currently going on in the sector, it is necessary not to give full control to my local devices with all the AI systems running in the same place.

For the LLM model, due to limited device specification resources, I had to use a self-hosted Llama.cpp running Qwen3.5 30b a3b. I also use the third-party services of OpenRouter. Furthermore, I decided to use two different AI agentic platforms, OpenClaw and Hermes Agent, to experiment with their advantages and disadvantages.

I use OpenClaw for slightly more private and personal interactions, complemented with the local LLM of the aforementioned models running on Llama.cpp. On Hermes Agent, I use the free routing of OpenRouter.

One of the current challenges or limitations in the agentic sector is about memory and context. All conversations are practically limited by the current conversation within a session. Each platform deals with these challenges differently, and yet, AFAIK, they are still exploring the most suitable design. This is why I self-hosted Honcho Memory. So far, it is able to retain my context across all conversations.

What it does: I am still exploring more, but it currently does these main jobs:

  • Managing my financial records/txs.
  • Basic researching skills.
  • Daily news briefing.
  • Personal calendar/schedule management.

It is fascinating how it does all of those automatically, given, to some degree, how tedious and repetitive the workflows or processes are. So it slightly boosted my productivity.

I'm excited for the way ahead! Will share how I explore it further and what it does to my workflow processes.