Two distinct components define any AI tool:
- The Model (“brain in a jar”): The LLM itself — determines how smart the AI is.
- The Harness (“the body/hands”): Everything else — how the AI accesses files, remembers sessions, uses tools, and fits your workflow. Determines how useful the AI is.
Model quality is converging across providers. Harness quality is diverging. The strategic question is no longer “which AI is smarter?” but “which harness fits how I actually work?”
Case comparison:
- Claude Code: “Collaborator at your desk” — full local shell access, Unix primitives, structured artifact memory.
- Codex: “Contractor in a clean room” — sandboxed cloud containers, no local access, codebase as sole system of record.
Neither is universally better. They encode different philosophies about trust, risk, and workflow.