Why final answers are not enough
Traditional evaluations often reduce an agent's performance to whether its final answer is correct. That metric is useful, but it hides important differences in how the answer was produced.
Different paths to the same result
Two agents can reach the same result while one uses unnecessary tools, makes unsupported assumptions, or takes a much longer path to completion.
What we measure instead
We are exploring evaluation frameworks that measure reliability, efficiency, evidence usage, tool selection, recovery behavior, and consistency alongside final task accuracy.