A model can reason. Capability is not a system.

In my view, LLMs are becoming a commodity, so the advantage comes from how you use the model.

A model alone can reason, yet it can't say what happened five minutes ago, which tool it may use, where work should go, or how it will prove the result.

The agent harness is the system around the model, with seven jobs: instructions, context, state, tools, orchestration, permissions, and traces plus evals.

A better agent needs work on both the model and the harness.