The central argument is that coding-agent harnesses with short system prompts interfere less with models they were not designed around, reduce per-request overhead, and expose more of each model's native behavior.
A reported 30-task comparison found that one Pi-derived harness passed 17 tasks, Claude Code and Codex passed 16 each, and OpenCode passed 14, while different harnesses led on speed and cost. The narrow test does not establish universal superiority, but it illustrates that harness design can materially affect results from the same model.
The guided tour highlights model and provider switching while retaining context, tree-based session branching, on-demand skills, TypeScript extensions, optional packages, and headless integration. Together, these features favor a small core with opt-in complexity.
Watch on YouTube
