StructAgent improves long-horizon digital agent execution via unified causal structure

Researchers present StructAgent, a state-centered framework designed to address the challenges of long-horizon computer-use tasks by explicitly structuring agent state and workflow around a unified causal representation of task progress.

• Introduces a unified state for maintaining compact, verifiable task progress and a structured workflow regulated by verifier-backed state transitions.
• Enables explicit progress checkpointing, evidence-driven task completion, targeted failure recovery, and tool-supported execution grounded in verification.
• Improves Qwen3.5-9B success rate on OSWorld-Verified from 27.0% to 46.9% and Qwen3.5-27B from 31.6% to 62.2%.
• Achieves a new open-source state of the art of 78.9% with MiniMax-M3 on OSWorld-Verified.
• Demonstrates generalization beyond desktop environments to Minecraft tasks.

The framework consistently improves performance across a wide range of LLM and VLM backbones, offering a reliable method for interpreting, verifying, and recovering from complex task executions.