StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure

Digital agents often fail at long-horizon tasks because they rely on raw interaction histories that accumulate noise, failed attempts, and ambiguous progress. This paper asks whether explicitly structuring an agent's state and workflow—rather than treating history as a flat sequence—can improve reliability, interpretability, and task success rates in complex computer-use environments.

The authors introduce StructAgent, a framework centered on two core components: a unified state and a structured workflow. The unified state acts as a shared interface for planning, acting, and verification, storing only the minimum sufficient information: current requirements, useful values (e.g., file paths), and verified evidence. The structured workflow enforces a rigorous loop where the agent proposes subgoals, executes actions, and—crucially—only commits progress to the state if a verifier confirms the outcome. This design enables explicit checkpointing, targeted failure recovery, and evidence-driven task completion.

StructAgent consistently improves performance across various LLM and VLM backbones. On the OSWorld-Verified benchmark, it significantly boosted the success rates of Qwen3.5 models (e.g., from 31.6% to 62.2% for the 27B variant) and achieved a new open-source state-of-the-art result of 78.9% using the MiniMax-M3 model. The framework also demonstrates strong generalization capabilities, showing that the same state-centered logic can be applied to web navigation tasks and Minecraft, where the verification source is swapped from desktop GUI probes to inventory-based checks.

By moving away from raw history processing, StructAgent provides a more robust, auditable, and controllable architecture for autonomous agents. It demonstrates that long-horizon success is less about model scale alone and more about how effectively an agent can maintain a causal, verifiable representation of its own progress. This approach offers a path toward building agents that can reliably handle multi-step, real-world workflows that would otherwise cause standard models to lose track of their objectives.