StructAgent: Harnessing Long-Horizon Digital Agents with Unified Causal Structure

StructAgent introduces a unified causal representation for long-horizon computer-use agents. Instead of relying mainly on raw interaction histories, it maintains a compact, verifiable task state and regulates execution through verifier-backed state transitions. The framework supports explicit progress checkpoints, evidence-based completion, targeted recovery from failures, and tool-assisted execution. According to the paper’s abstract, OSWorld-Verified success rates rise from 27.0% to 46.9% for Qwen3.5-9B and from 31.6% to 62.2% for Qwen3.5-27B. With MiniMax-M3, it reaches 78.9% and also generalizes to Minecraft.