AlphaDiana, described in a new arXiv preprint, is a measurement system designed for agents operating on verifiable reasoning tasks. The paper frames agents as reasoning models paired with harnesses—the surrounding infrastructure that organizes model calls, executes tools, loads skill instructions, and retains history in memory.

According to the abstract, understanding agent performance requires linking execution conditions to outcomes. AlphaDiana appears to be built around that link, though the full methodology is not visible in the excerpt.

The work highlights the harness as a central component of agent behavior, suggesting that performance depends on more than just the reasoning model itself.