A new arXiv paper, CUAWright, takes aim at how computer-use agents are typically built. The authors note that the prevailing approach couples a model with a domain-specific harness—a browser or desktop environment equipped with human-engineered tools that are fixed before task execution. This design, they argue, is the default but may not be optimal.
The paper's proposed alternative is a minimal unified interface for digital agents. Rather than relying on bespoke, pre-configured tools, CUAWright seems to lean on models' coding capabilities to interact with digital environments. The abstract cuts off mid-sentence at "As models' coding…," so the full details of the mechanism are not available from the source.
Because only the opening lines of the abstract are provided, the digest is necessarily limited. The core contrast is clear: domain-specific harnesses versus a minimal, unified interface. But the specifics of how CUAWright implements that interface, and how it compares empirically to existing approaches, are not described in the available text.