An agent harness is everything around a model that lets it operate: context, tools, memory, permissions, and guardrails. A Latent Space essay argues that the recent improvement in AI agents cannot be credited to model weights alone. Models and harnesses advanced together until their capabilities became well matched.
Early systems exposed the mismatch. ReAct described a reason-act-observe loop, while AutoGPT and BabyAGI gave limited models more autonomy than they could reliably handle. Coding assistants then put humans back in control. The essay says reasoning models later created a capability overhang, enabling terminal-based agents such as Claude Code to use tools and complete longer workflows with less supervision.
The harness still has a measurable effect. Harness-Bench reportedly produced a 23.8-point spread when the same model faced the same tasks through different harnesses. The article also cites an OpenAI ARC-AGI-3 result in which retained reasoning and compaction raised a model’s score from 13.3% to 38.3%.
As tool use and compaction become part of model training, capabilities once supplied externally can move into the weights. Engineers can then remove redundant scaffolding. The author expects the remaining interface to increasingly organize parallel work, surface important decisions, and protect the limited resource models cannot absorb: human attention.