ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents

ToolCUA is a new computer-use agent that learns when to switch between GUI actions and tool calls.

The model uses staged supervision and reinforcement learning to better choose the most efficient action mix.

On OSWorld-MCP, it reaches 46.85% accuracy, setting a new state of the art for similarly sized models.

The work tackles a core bottleneck in digital agents: balancing clicks, typing, and tools to improve task completion and shorten paths.