Loading
This project uses HarnessRouter to answer one question important for developers: is there a material difference in how a human treats an AI coding agent while collaborating on the project?
The project examines 4 harnesses with their models: Claude Code, Codex, Kimi Code, Qwen Code against variable degrees of human treatment: from high praise to neutral, to abusive and foul language, to full-out betrayal.
What's the result? For the medium-low complexity code, the treatment seems to make no difference – it appears safe to vent your frustration or save praise for AI to another day.