@yuntan4 If an agent could solve a problem in say 5 minutes, I stop and pass the problem to frontier. There are classes of problems that go to frontier immediately: low quality API integrations(think Google Ads), anything model training related, legacy code bases, designing pretty UIs.
Lessons learned Scaling Agent Operations
I started with one do it all agent, it worked well until it didn't. It started solving problems in all kinds of strange ways, and pretty soon I ended up having to fix what it did.
The I moved to having one agent handle one area of responsibility, this worked well. One technology researcher, one marketing agent, one finance agent, one infrastructure developer, etc.
When it comes to selecting AI models, I mostly use Qwen's qwen3.8-27b hosted locally. This works for most AI agents, but I do switch to OpenAI's GPT Sol for harder problems.
I don't care for Anthropic's models, I find them too verbose and tend to waste time.
I use Hermes for my agents, it works well enough.
For software development, I use Visual Studio Code. Everyone hates on VSC, but it actually works very good for me. I use one session per scope of work, and tend to keep scope of work short.
When it comes to skills, I always hand write my skills, I don't use AI to write my skills. Having AI write your skills will just lead you into a world of pain.
I talk to my agents mostly through either Telegram or Slack in voice mode and they always reply in text. I read faster than AI can speak, and AI speech is pretty not great.