Large language model (LLM) agents are increasingly deployed on long-horizon tasks that involve using tools and modifying real workspaces. According to the paper, how well these agents perform depends as much on the harness around the model as on the model itself.

To study this, the paper introduces Doc2Harness, a benchmark for evaluating LLMs on harness engineering for coding agents. The benchmark focuses on the construction of the surrounding infrastructure—the harness—that enables agents to operate over extended tasks.

The abstract suggests that progress in agent performance may come as much from improving harness engineering as from improving the models. By measuring this capability, Doc2Harness aims to draw attention to a component of agent development that is often overlooked.