After testing 50+ personal agents, harness matters more than the model
The author tested 50+ personal agents including Grok Bot, Dots, Muse and Instinct, finding most fail on system prompts and tool definitions—some even think they are Claude Code and don't know their own routines.
I've tested 50+ Grok Bot, Dots, Muse, Instinct, alternatives. Its true, see my site on Bio. Most of them fail at 1 basic point: the system prompt, the docs, and the harness commands... and bla bla bla. It doesn't matter if you've built a Ferrari if it thinks it's a turtle on the inside. Real example: in several of these apps, I asked the agent to create a routine for me. It did... inside Claude. INSIDE CLAUDEEEE?! 🤦♂️ The agent didn't even know it had its own routines system. It thought it was Claude Code. The agent has to know who it is. Even running on Claude, it isn't Claude Code. It's Software X, with Software X's tools, limits and commands. If you're building a harness today, study the leaked system prompts from Grok Bot, Dots, Pi, OMP, OpenCode... That's the bare minimum. There's no point asking Claude to "build a better Grok Bot" if you don't understand the basic structure of a harness. A great model can't save a bad harness. Pro tip: if you're digging into these system prompts, DON'T use Claude, Codex, Grok or Gemini for it. Go with GLM, Kimi or DeepSeek. Nobody wants to get banned 😂 Which alternative disappointed you the most, and which one actually surprised you? 👇