A new research paper, "From Model Scaling to System Scaling: Scaling the Harness in Agentic AI," posits that stronger AI agents will emerge from better surrounding systems, termed the 'harness,' rather than solely from larger models. This harness manages what the model perceives, its tool usage, memory, and action verification, with key areas for improvement being context control, trustworthy memory, and routing to tools or helper agents.
Agentic AI deployment is bottlenecked by system design and reliability, not just model capability, shifting focus to infrastructure for context, memory, and tool orchestration.