My minimalistic AI Agent setup
I decided to just put Claude Code in a docker container as a my very minimal AI agent.
I simply defined a Dockerfile and a docker-compose.yml that give Claude Code everything it needs to help me write code that will be pushed onto GitHub for me to review.
I run 3 docker containers that are exactly the same with a session of Claude Code on my linux server. I share how I set up my linux server in this article.
I use SSH via iTerm2 and tmux to access my Claude Code agents to tell them to write code and push PRs onto GitHub for me to review.
Why I’m Building My Own
I know that Open Claw and Hermes is all the rage right now BUT, I decided to build my first AI Agents myself.
My agents are very basic. They are focused on being good software developers that increase the quality and quantity of useable apps that I can publish.
The reason is very simple.
I don’t need any of the features that these frameworks come with right now… because I don’t know how to use them, yet.
I think it’s a mistake ask your brain to juggle too many things when you’re learning something new.
The main tasks I need my AI agent to work on are coding tasks so I’m going to deep dive on setting up AI agents that can help me build apps as a first priority.
I find Claude Code already does this really well so I decided to just containerize with Docker containers and spin up multiple sessions of Claude Code so that they can work on coding tasks in parallel.
Hidden Cost of Running OpenClaw and Hermes
It might sound great to have an AI assistant that can be proactively doing all the things for you…
But lets be honest, it’s likely to be another thing for you to babysit.
OpenClaw will try to reorganize all your folders and you’ll hate it so you spend a day reverting it back to how you prefer things to be organized.
Hermes might optimize your calendar but then you find out that you don’t like the schedule it’s created and you spend an afternoon rearranging your meeting times.
Of course, the ideal case is that it’ll do the work that you didn’t want to do.
But the reality is that it’ll do all sorts of things that you never wanted to be done.
And you’ll have to fix the things it broke. That takes time and mental energy.
So instead of handing the keys to the kingdom to an AI that you don’t know how to work with, you can build an ai partner that actually saves you time by refining one task at a time.
By slowly handing off tasks and training them to be good at that tasks before delegating another you’ll save yourself the headache of fighting multiple fires that the AI agents will inevitably start when they first take on a new task.
