Both of them expressed the importance of "getting out of [their] comfort zone", despite age.
Anthropic 点名的三家公司，抓取的重心都落在代理行为 (agentic behavior) 这同一个方向上，包括 AI 自主规划、工具调用、分解复杂任务并逐步执行的能力等。
，详情可参考Line官方版本下载
Earlier this week he told Variety: "I remember there was a microphone just in front of me, and with hindsight I have to question whether this was wise, so close to where I was seated, knowing I would tic."
为此，智能体工程通常通过以下几类手段给大模型加“外骨骼”以改善可靠性：引入检索与知识库（RAG）以降低幻觉和知识陈旧的影响；预先设计和约束工作流，而不是完全自由的“自治智能体”，以此限定可接受的执行路径；通过多次回答、自一致性检查或模型间交叉验证，识别并过滤高风险输出；在关键链路节点上设置人工审批，让人类对高风险动作“最后拍板”。
圖像來源，GREG BAKER/AFP via Getty Images