Much office work is investigation: the answer exists, but hidden in messy files and disconnected systems. Large language models (LLMs), text-generating AI, are usually tested on making documents, not finding evidence. A harness, the shell and filesystem that lets a model write and run programs, keeps it working through long tasks.
A new ran four harnesses on the same mid-tier model with real tax extracts and a hidden correct answer. Two model-neutral harnesses recovered about half the score. The two lab harnesses recovered nothing with another model.
Top-tier models solved the same tasks in about twenty-five minutes. Below that tier, the harness decides how far a model gets, so generic agent infrastructure can matter as much as model quality.