From Model Scaling to System Scaling: Scaling the Harness in Agentic AI

A paper argues that agentic AI should be advanced by “scaling the harness” around foundation models, not only by making models bigger.

The harness includes modular pieces such as memory, skill routing, orchestration, verification, and context governance.

The authors present CheetahClaws as a reference harness to support comparison, benchmarking, and more auditable system design.

The broader point: better architecture, evaluation, and governance may matter as much as model size for agentic AI progress.