What I use AI for
- CUDA kernel optimization
- Other verifiable tasks where the agent can check its work
My setup
I built my own harness for maximum control over what my agent does: https://github.com/99991/guppycode
- Tool calls are executed safely in a stateless Docker sandbox
- Minimal tools for minimal token use
- No third-party dependencies to avoid supply chain attacks
- Less than 1000 lines of Python (could be less than 100 without all the bells and whistles)
I currently use the $10 per month OpenCode Go subscription, but you can also use DeepSeek directly or through OpenRouter for around the same price.
Advice
- Build your own harness. It is not difficult and will give you a much deeper understanding of AI agents. This is a good starting point (less than 100 lines): https://github.com/99991/MinimalAgent
- Most skills are snake oil and should be replaced with proper documentation.