许多人认为智能体是人工智能的终极目标。Stuart Russell 和 Peter Norvig 的经典著作《人工智能：一种现代方法》（Prentice Hall, 1995）将 AI 研究领域定义为“对理性智能体的研究和设计”。基础模型前所未有的能力为之前难以想象的智能体应用打开了大门。这些新能力使得开发自主、智能的智能体成为可能，它们可以作为我们的助手、同事和教练。它们可以帮助我们创建网站、收集数据、规划旅行、进行市场调研、管理客户账户、自动化数据录入、为我们准备面试、面试候选人、谈判交易等。可能性似乎是无限的，这些智能体的潜在经济价值巨大。本节将从智能体的概述开始，然后继续讨论决定智能体能力的两个方面：工具和规划。智能体以其新的操作模式，也有新的失败模式。本节将以如何评估智能体以捕捉这些失败的讨论结束。
Intelligent agents are considered by many to be the ultimate goal of AI. The classic book by Stuart Russell and Peter Norvig, _Artificial Intelligence: A Modern Approach_ (Prentice Hall, 1995), defines the field of AI research as “_the study and design of rational agents._” The unprecedented capabilities of foundation models have opened the door to agentic applications that were previously unimaginable. These new capabilities make it finally possible to develop autonomous, intelligent agents to act as our assistants, coworkers, and coaches. They can help us create a website, gather data, plan a trip, do market research, manage a customer account, automate data entry, prepare us for interviews, interview our candidates, negotiate a deal, etc. The possibilities seem endless, and the potential economic value of these agents is enormous. This section will start with an overview of agents and then continue with two aspects that determine the capabilities of an agent: tools and planning. Agents, with their new modes of operations, have new modes of failure. This section will end with a discussion on how to evaluate agents to catch these failures.
核心贡献 · Key contributions
将 AI 智能体定义为通过传感器感知环境、通过执行器采取行动的实体，并以基础模型作为其大脑。 Defines AI agents as entities perceiving environment via sensors and acting via actuators, with foundation models as the brain.
将规划与执行解耦：先生成计划，验证后再执行，减少无效的 API 调用。 Decouples planning from execution: generate plan, validate, then execute, reducing fruitless API calls.
展示工具使用显著提升模型性能；Chameleon 在 ScienceQA 上提升 11.37%，在 TabMWP 上提升 17%。 Shows tool use significantly boosts model performance; Chameleon improves ScienceQA by 11.37% and TabMWP by 17%.
引入 ReAct 和 Reflexion 模式：将推理与行动交错进行，并通过自我反思进行错误纠正。 Introduces ReAct and Reflexion patterns: interleaving reasoning and action with self-reflection for error correction.
讨论规划失败：工具误用、目标失败和反思错误；提出评估指标如计划有效比率。 Discusses planning failures: tool misuse, goal failure, and reflection errors; proposes evaluation metrics like plan validity rate.
强调由于复合错误和更高风险，智能体需要更强大的模型。 Highlights that agents require more powerful models due to compound mistakes and higher stakes.
局限 · Limitations
自回归模型可能从根本上缺乏规划能力，因为无法回溯，正如 Yann LeCun 所论证的。 Autoregressive models may fundamentally lack planning ability due to inability to backtrack, as argued by Yann LeCun.
智能体性能随步骤增加而下降；每步 95%的准确率在 10 步后降至 60%。 Agent performance degrades with more steps; 95% per-step accuracy drops to 60% over 10 steps.
工具选择具有挑战性；更多工具增加能力但也增加复杂性和上下文窗口限制。 Tool selection is challenging; more tools increase capability but also complexity and context window constraints.
智能体在工具调用中容易产生幻觉，生成无效函数或错误参数。 Agents are prone to hallucination in tool calls, generating invalid functions or wrong parameters.
反思和错误纠正因额外生成 token 而增加延迟和成本。 Reflection and error correction add latency and cost due to extra token generation.