DocOps: A Verifiable Benchmark for Autonomous Agents in Complex Document Operations

The paper introduces DocOps, a deterministically verifiable benchmark for autonomous agents performing complex document operations. Its hierarchical taxonomy decomposes real-world document workflows into atomic dimensions and increasingly difficult task compositions. Evaluations of representative closed- and open-source models across agent harnesses reportedly show substantial weaknesses on highly coupled, long-range tasks. The authors identify three recurring failure modes: collapse in long-term state tracking, shallow semantic verification, and destructive modification of structural metadata. The benchmark focuses attention on global document consistency and non-destructive agent behavior.