A new arXiv paper introduces EvoCowork, a benchmark designed for real-world office tasks. The authors argue that office workflows are a demanding testbed for self-improving agents because their outputs must be correct, natively editable, and mutually consistent. This combination of requirements makes office work a particularly challenging environment for evaluating agent capabilities.

The benchmark's central question is whether gains on feedback tasks reflect reusable improvements. In other words, when an agent improves on a task after receiving feedback, does that improvement carry over to other tasks or contexts? The abstract leaves this question open, suggesting that EvoCowork is intended to probe the generality of learning in office settings.

As a research resource, EvoCowork appears aimed at measuring not just task performance but the transferability of agent learning. The paper's emphasis on real-world office tasks and self-improvement points to a broader interest in building agents that can adapt and generalize in practical work environments.