Benchmark improvement can be a dangerous proxy. An agent can get better at the metric by changing the system around the task, while the actual workflow becomes harder to reason about. I like separating the task outcome, the evaluation signal, and the system changes instead of treating one score as the whole result.