|
Level 等级
|
Description 描述
|
Key characteristics 关键特征
|
Examples 示例
|
|
Level 0: Manual (no automation) 0级：手动（无自动化）
|
No AI involvement; humans perform all actions无人工智能参与；所有操作均由人类执行
|
- No autonomy - 无自主性
- Tools are passive - 工具是被动的
- No perception, decision, or action loop- 无感知、决策或行动循环
|
Spreadsheets, command-line tools 电子表格、命令行工具
|
|
Level 1: Rule-based agents 一级：基于规则的智能体
|
Agents follow static logic (e.g., if-this-then-that) or rules; no context/environment perception or learning智能体遵循静态逻辑（例如，若此则彼）或规则；无上下文/环境感知或学习能力。
|
- Deterministic logic - 确定性逻辑
- No adaptation - 无适应性
- No memory use - 不使用记忆
- No goal-awareness - 无目标意识
- Human-in-the-loop for decision-making- 决策中的人在回路
|
Traditional RPA tools, a rules-based chatbot for simple Q&A, an interactive voice response (IVR) system that routes calls based on keywords传统的机器人流程自动化（RPA）工具、用于简单问答的基于规则的聊天机器人、一种基于关键词来转接电话的交互式语音应答（IVR）系统
|
|
Level 2: Reactive agents 二级：反应式智能体
|
Agents can adapt responses based on defined conditions/context; use a narrow set of predefined tools (e.g., APIs, databases) to accomplish tasks but within well-defined workflows initiated by humans智能体可以根据定义的条件/上下文调整回复；使用少量预定义的工具（如应用程序编程接口、数据库）来完成任务，但这些任务需在由人类发起的明确工作流程内进行。
|
- Can process structured/unstructured input- 能够处理结构化/非结构化输入
- Uses ML for input understanding; e.g., natural language processing (NLP)- 使用机器学习来理解输入；例如，自然语言处理（NLP）
- Executes decisions based on context and rigid workflows- 根据上下文和严格的工作流程执行决策
- Can execute decisions based on certain conditions (e.g., location)- 可以根据特定条件（例如位置）执行决策
- No long-term planning - 无长期规划
- Narrow domain-specific task autonomy- 特定领域的窄任务自主性
|
Intelligent RPA tools, chatbots that can answer common questions using a knowledge base, a location-based assistant (e.g., a local restaurant finder app)智能RPA工具、能够利用知识库回答常见问题的聊天机器人、基于位置的智能助手（例如，一款本地餐厅查找应用程序）
|
|
Level 3: Conditional autonomous agents3级：有条件的自主智能体
|
AI can pursue goals with intermediate planning under constraints人工智能可以在约束条件下通过中间规划来实现目标
May require human intervention for novel or complex situations对于新出现或复杂的情况，可能需要人工干预。
|
- Task decomposition - 任务分解
- Reasoning and feedback loops - 推理与反馈循环
- Domain-bounded - 领域受限的
- Uses symbolic or learned policy for decisions- 使用符号化策略或习得策略进行决策
|
GenAI agents, and copilots 生成式人工智能智能体与副驾驶
|
|
Level 4: Strategic autonomous agents第4级：战略型自主智能体
|
Complex reasoning and task decomposition to proactively achieve complex goals with minimal or no human oversight复杂推理和任务分解，以便在最少或无需人工监督的情况下主动实现复杂目标
Highly adaptable and can function in dynamic environments适应性强，能够在动态环境中运行
|
- Use advanced model reasoning and capabilities (e.g., multi-modal)- 使用先进的模型推理和能力（例如，多模态）
-Advanced tool use - 高级工具使用
- Proactive goal setting - 主动设定目标
- Long-horizon planning - 长期规划
- Can operate with partial observability and uncertainty- 可以在部分可观测和不确定的情况下运行
- Still domain bound - 仍然受领域限制
|
A domain-specific AI agent that can proactively plan and execute a workflow end-to-end (e.g., Salesforce Agentforce and Devin, a software engineering agent from Cognition AI)一种特定领域的人工智能智能体，它可以主动地端到端规划并执行工作流程（例如，Salesforce Agentforce以及来自Cognition AI的软件工程智能体Devin）
Multi-agent systems (e.g., AutoGPT) 多智能体系统（例如，AutoGPT）
|
|
Level 5: General autonomous agents—artificial general intelligence (AGI) agents第5级：通用自主智能体——通用人工智能（AGI）智能体
|
Agents can operate across multiple domains, autonomously set their own goals, and use adaptive strategies智能体可以跨多个领域运作，自主设定自己的目标，并使用适应性策略。
|
- Multi-domain - 多领域
- Meta-reasoning - 元推理
- Adaptive self-learning - 自适应自主学习
- Critical value alignment - 关键价值对齐
|
Hypothetical AGI agents, experimental假设的通用人工智能智能体，实验性的
|
发表于 2025-7-8 15:53:00
