AI coding tools depend on the harness, not just the model
Tools like Claude Code do not work through the alone. The creates tokens and signals when a tool should be used.
The harness reads those tool-use signals, runs the actual action, and sends the result back to the model. This means many AI agent scores are really measuring the model and the harness together, not the model by itself.
The harness is where , checks, retries, and oracles are handled. Anthropic, OpenAI, and other AI companies are competing not only on better models, but also on the surrounding system that makes those models useful in real work.
Key points
- A does not directly use tools; it creates signals for tool use.
- The harness interprets those signals, runs actions, and feeds results back.
- AI agent scores often reflect both the model and the harness together.
- , checks, retries, and oracles live in the harness.
- The practical gap between can come from the surrounding system, not only the model.