CUAWright: A Minimal Unified Interface for Digital Agents
The paper argues that agents do better when their only interface is a programmable terminal, not a prebuilt GUI harness.
CUAWright is described as a roughly 3,000-line terminal harness where agents act through bash commands and use the file system to create tools and manage context. The authors report stronger results than GUI-native or hybrid CLI setups across OSWorld 2.0, Online-Mind2Web, Odysseys, CADGenBench, and BenchCAD. On OSWorld 2.0, they claim a 33.2% relative gain in partial reward and a 37.5% estimated cost reduction versus the published GPT-5.5 baseline. The stated implication is that many digital environments are more directly programmable than their interfaces suggest. HF Daily Papers' note
CUAWright is described as a roughly 3,000-line terminal harness where agents act through bash commands and use the file system to create tools and manage context. The authors report stronger results than GUI-native or hybrid CLI setups across OSWorld 2.0, Online-Mind2Web, Odysseys, CADGenBench, and BenchCAD. On OSWorld 2.0, they claim a 33.2% relative gain in partial reward and a 37.5% estimated cost reduction versus the published GPT-5.5 baseline. The stated implication is that many digital environments are more directly programmable than their interfaces suggest. HF Daily Papers' note
score 5