📄 中文摘要
最佳的智能体评估直接衡量我们关心的智能体行为。通过精心策划的数据来源、创建有效的指标以及长期进行有针对性的实验，提升智能体的准确性和可靠性。评估不仅是数据的收集，更是对智能体行为的塑造。通过不断的评估和反馈，智能体能够在特定任务中表现得更加出色，从而满足实际应用需求。评估过程中的每一步都至关重要，确保最终结果能够反映智能体的真实能力和潜力。
我们如何为深度智能体构建评估
出处: How we build evals for Deep Agents
发布: 2026年3月26日
📄 English Summary
How we build evals for Deep Agents
The best evaluations for agents directly measure behaviors that are of interest. By carefully sourcing data, creating effective metrics, and conducting well-scoped, targeted experiments over time, the accuracy and reliability of agents can be improved. Evaluations are not just about data collection; they shape agent behavior. Continuous assessment and feedback enable agents to perform better in specific tasks, meeting real-world application needs. Each step in the evaluation process is crucial to ensure that the final results accurately reflect the true capabilities and potential of the agents.