AlphaDiana is a measurement framework for evaluating agents that perform verifiable reasoning tasks. It offers modular interfaces for datasets, agents, environments, and scorers, preserving the native harness control loop while independently managing file reuse and memory retention.\ \ The system supports two analysis modes: macro‑level comparisons of complete agents, and micro‑level interventions on components such as tools, skill instructions, and memory under matched conditions. Experiments involved three open‑source models and three open‑source harnesses across five answer‑based benchmarks; two models were also tested on two interactive benchmarks.\ \ Findings reveal that harnesses do not consistently improve reasoning accuracy: gains on questions unsolvable by the model alone can be offset by losses on questions the model answers correctly without tools. The same inconsistency appears in component interventions—neither skill guidance nor retained history reliably boosts accuracy.\ \ Behavior and output‑token usage must be assessed together with correctness: low token entropy may accompany wrong answers, and fewer tokens per execution does not guarantee lower overall consumption—when accuracy drops, tokens per correct answer can actually increase. AlphaDiana links answer correctness and execution status to trace logs and resource records, enabling fair comparison of agents and diagnosis of failures under shared task and resource policies.\ \ Review