Agent evaluation now has to measure the whole system
Princeton’s SAgE survey maps 33 trends for evaluating agents as systems, so scores need to account for the model’s scaffold, tools, testing environment, and user interface, not just the model. It organizes the open questions around how evaluations work, what they measure, and who conducts them, with concrete distinctions such as pass@k (any of k runs succeeds) versus pass∧k (all k runs succeed).