Recent advancements in large language models (LLMs) and vision-language models (VLMs) have enabled increasingly capable digital agents for computer use. However, real-world tasks are often long-horizon, involving evolving contexts with accumulated observations, intermediate edits, failed attempts, and partially completed executions.
Existing agents typically operate over raw interaction history, making task progress difficult to interpret, verify, and recover, ultimately limiting reliable long-horizon execution. This paper argues that addressing this challenge requires explicitly structuring both the agent's state and workflow around a unified causal representation of task progress.
We present StructAgent, a state-centered framework that maintains compact, verifiable task progress and regulates it through verifier-backed state transitions. Building on this design, StructAgent enables explicit progress checkpointing, evidence-driven task completion, targeted failure recovery, and tool-supported execution, ensuring that all progress updates remain grounded in verification.
Extensive experiments demonstrate that StructAgent consistently improves a wide range of LLM and VLM backbones on long-horizon computer-use tasks. On OSWorld-Verified, it improves Qwen3.5-9B from 27.0% to 46.9% success rate and Qwen3.5-27B from 31.6% to 62.2%, achieving a new open-source state of the art of 78.9% with MiniMax-M3. Moreover, the same framework generalizes beyond desktop environments to Minecraft, showcasing the generality of our design.
Blogger's Review: The introduction of StructAgent offers a novel approach to executing long-horizon tasks by enhancing the reliability and verifiability of digital agents through a unified causal structure. This method not only improves the performance of existing models but also demonstrates its applicability across various environments, providing significant research value and potential for real-world applications.