The harness around a model decided more of the benchmark score than the model itself. So treat scaffold selection as a first-order engineering decision.

This survey pairs its synthesis with matched-task diagnostics, and finds system differences that are measurable while the model-variant differences are not.

A stronger model variant inside the same harness bought no extra resolved tasks, only latency.

– arxiv. org/abs/2608.20485

Title: "Terminal Agents: A Survey of AI Agents in Command-Line Environments"