🤖 AI不是独行侠：人在回路模式让机器学会"求助"
最近和一位做自动驾驶的朋友聊天，他告诉我一个有趣的现象：即使在技术最先进的无人车上，遇到复杂路况时，系统还是会"犹豫"，然后请求远程人工协助。
这让我想起了一个重要概念：人在回路（Human-in-the-Loop）[HITL]。
说白了，这就是让AI学会"求人"的艺术。
📌 什么是HITL？不是监督，是协作
很多人以为HITL就是人类监督AI工作，其实远不止如此。
HITL的核心是建立人机协同生态系统。在这个系统里，AI负责处理海量数据和重复性任务，人类则专注于需要判断力、创造力和道德决策的环节。
举个简单例子：医疗AI可以分析X光片，发现可疑病灶，但最终诊断报告必须由医生签字确认。这不是因为AI不够聪明，而是因为生命攸关的决策需要人类承担责任。
⚠️ 为什么完全自动化不一定是好事？
完全自主的AI系统存在天然缺陷：
首先，AI缺乏常识推理能力。它可能在图像识别上超越人类，但面对前所未见的情况时容易"犯傻"。
其次，AI无法理解伦理边界。一个自动驾驶系统可能为了避让行人而选择撞墙，但如果有乘客呢？这个道德困境需要人类判断。
最后，责任归属问题。当AI做出错误决策时，谁承担责任？HITL通过明确的人类介入点，让责任链条清晰可见。
✅ HITL的三大核心价值
安全性保障是最直接的价值。在金融交易中，大额转账需要人工审批；在内容审核中，AI先过滤明显违规内容，人类处理边界案例。
准确性提升体现在专业领域。法律文书审查AI可以快速筛选相关条款，但律师需要验证AI的发现是否准确，是否适用于具体案情。
责任明确化让各方都有安全感。医生使用AI辅助诊断，但最终诊断书由医生签署；程序员用AI生成代码，但需要人工审查和测试。
🎯 HITL不是万能的，但很有用
说实话，HITL也有明显的局限性。
可扩展性是个大问题。人类监督者无法处理数百万级别的任务，这决定了HITL更适合高风险、低频次的场景，而不是日常简单任务。
专业依赖性很强。一个不懂编程的人很难发现AI生成代码中的微妙错误。这要求参与HITL的人类必须具备相应的专业知识。
隐私考量不容忽视。敏感信息在暴露给人类操作员前需要匿名化处理，这增加了系统的复杂性。
💡 实际应用：从理论到实践
内容审核是最常见的应用场景。AI先过滤掉明显的违规内容，人类审核员处理那些需要语境理解的边界案例。
金融风控中，AI监控交易模式，发现异常后提交给人类分析师做最终判断。这样既保证了效率，又确保了准确性。
技术支持领域，AI处理常见问题，复杂问题升级到人工客服。很多公司的客服系统都在用这种混合模式。
自动驾驶可能是最直观的例子。车辆可以自主处理大部分驾驶任务，但在极端天气或复杂路况下会请求远程人工接管。
🔍 写在最后
HITL给我们的最大启示是：最有效的AI系统不是完全自动化的，而是能够巧妙结合人类判断力与机器计算力的混合系统。
在实际工作中，我们需要根据业务场景的风险程度和复杂性，设计恰当的HITL介入点。既保持AI的效率优势，又确保关键决策的人类监督。
毕竟，AI再智能，也需要人类来教它什么时候该"求人"。
你觉得在哪些场景下，AI应该主动寻求人类帮助？欢迎留言分享你的看法。
今天就试试！观察一下你身边的AI应用，看看它们是如何处理复杂情况的。
Chapter 13: Human-in-the-Loop
第13章：人在回路模式
The Human-in-the-Loop (HITL) pattern represents a pivotal strategy in the development and deployment of Agents. It deliberately interweaves the unique strengths of human cognition—such as judgment, creativity, and nuanced understanding—with the computational power and efficiency of AI. This strategic integration is not merely an option but often a necessity, especially as AI systems become increasingly embedded in critical decision-making processes.
人在回路（HITL）模式代表了智能体开发和部署中的关键策略。它有意地将人类认知的独特优势——如判断力、创造力和细微理解能力——与人工智能的计算能力和效率相结合。这种战略整合不仅仅是一种选择，通常更是一种必然，特别是当人工智能系统日益嵌入关键决策过程时。
The core principle of HITL is to ensure that AI operates within ethical boundaries, adheres to safety protocols, and achieves its objectives with optimal effectiveness. These concerns are particularly acute in domains characterized by complexity, ambiguity, or significant risk, where the implications of AI errors or misinterpretations can be substantial. In such scenarios, full autonomy—where AI systems function independently without any human intervention—may prove to be imprudent. HITL acknowledges this reality and emphasizes that even with rapidly advancing AI technologies, human oversight, strategic input, and collaborative interactions remain indispensable.
HITL的核心原则是确保人工智能在道德边界内运行，遵守安全协议，并以最佳效果实现其目标。这些关注点在以复杂性、模糊性或重大风险为特征的领域中尤为突出，在这些领域中，人工智能错误或误解的影响可能相当严重。在这种情况下，完全自主——即人工智能系统在没有任何人工干预的情况下独立运行——可能被证明是不明智的。HITL承认这一现实，并强调即使人工智能技术快速发展，人类监督、战略输入和协作互动仍然不可或缺。
The HITL approach fundamentally revolves around the idea of synergy between artificial and human intelligence. Rather than viewing AI as a replacement for human workers, HITL positions AI as a tool that augments and enhances human capabilities. This augmentation can take various forms, from automating routine tasks to providing data-driven insights that inform human decisions. The end goal is to create a collaborative ecosystem where both humans and AI Agents can leverage their distinct strengths to achieve outcomes that neither could accomplish alone.
HITL方法从根本上围绕着人工智能与人类智能协同的理念。HITL不是将人工智能视为人类工作者的替代品，而是将其定位为增强和提升人类能力的工具。这种增强可以采取多种形式，从自动化日常任务到提供数据驱动的见解来指导人类决策。最终目标是创建一个协作生态系统，在这个系统中，人类和人工智能智能体都可以利用各自的独特优势，实现任何一方单独都无法完成的成果。
In practice, HITL can be implemented in diverse ways. One common approach involves humans acting as validators or reviewers, examining AI outputs to ensure accuracy and identify potential errors. Another implementation involves humans actively guiding AI behavior, providing feedback or making corrections in real-time. In more complex setups, humans may collaborate with AI as partners, jointly solving problems or making decisions through interactive dialog or shared interfaces. Regardless of the specific implementation, the HITL pattern underscores the importance of maintaining human control and oversight, ensuring that AI systems remain aligned with human ethics, values, goals, and societal expectations.
在实践中，HITL可以通过多种方式实现。一种常见的方法是让人类充当验证者或审查者，检查人工智能输出以确保准确性并识别潜在错误。另一种实现方式是让人类主动指导人工智能行为，提供反馈或实时进行纠正。在更复杂的设置中，人类可以与人工智能作为合作伙伴进行协作，通过交互式对话或共享界面共同解决问题或做出决策。无论具体实现方式如何，HITL模式都强调了保持人类控制和监督的重要性，确保人工智能系统与人类的道德、价值观、目标和社会期望保持一致。
Human-in-the-Loop Pattern Overview
人在回路模式概述
The Human-in-the-Loop (HITL) pattern integrates artificial intelligence with human input to enhance Agent capabilities. This approach acknowledges that optimal AI performance frequently requires a combination of automated processing and human insight, especially in scenarios with high complexity or ethical considerations. Rather than replacing human input, HITL aims to augment human abilities by ensuring that critical judgments and decisions are informed by human understanding.
人在回路（HITL）模式将人工智能与人类输入相结合，以增强智能体能力。这种方法承认，最佳的人工智能性能通常需要自动化处理和人类洞察力的结合，特别是在具有高复杂性或伦理考量的场景中。HITL不是取代人类输入，而是通过确保关键判断和决策基于人类理解来增强人类能力。
HITL encompasses several key aspects: Human Oversight, which involves monitoring AI agent performance and output (e.g., via log reviews or real-time dashboards) to ensure adherence to guidelines and prevent undesirable outcomes. Intervention and Correction occurs when an AI agent encounters errors or ambiguous scenarios and may request human intervention; human operators can rectify errors, supply missing data, or guide the agent, which also informs future agent improvements. Human Feedback for Learning is collected and used to refine AI models, prominently in methodologies like reinforcement learning with human feedback, where human preferences directly influence the agent's learning trajectory. Decision Augmentation is where an AI agent provides analyses and recommendations to a human, who then makes the final decision, enhancing human decision-making through AI-generated insights rather than full autonomy. Human-Agent Collaboration is a cooperative interaction where humans and AI agents contribute their respective strengths; routine data processing may be handled by the agent, while creative problem-solving or complex negotiations are managed by the human. Finally, Escalation Policies are established protocols that dictate when and how an agent should escalate tasks to human operators, preventing errors in situations beyond the agent's capability.
HITL包含几个关键方面：人类监督，涉及监控人工智能智能体的性能和输出（例如，通过日志审查或实时仪表板），以确保遵守指南并防止不良结果。干预和纠正发生在人工智能智能体遇到错误或模糊场景并可能请求人工干预时；人类操作员可以纠正错误、提供缺失数据或指导智能体，这也为未来的智能体改进提供信息。用于学习的人类反馈被收集并用于改进人工智能模型，特别是在如带有人类反馈的强化学习等方法中，人类偏好直接影响智能体的学习轨迹。决策增强是指人工智能智能体向人类提供分析和建议，然后由人类做出最终决定，通过人工智能生成的见解而非完全自主来增强人类决策。人机协作是人类和人工智能智能体贡献各自优势的合作互动；常规数据处理可能由智能体处理，而创造性问题解决或复杂谈判则由人类管理。最后，升级策略是建立的协议，规定智能体应在何时以及如何将任务升级给人类操作员，防止在智能体能力范围之外的情况下出现错误。
Implementing HITL patterns enables the use of Agents in sensitive sectors where full autonomy is not feasible or permitted. It also provides a mechanism for ongoing improvement through feedback loops. For example, in finance, the final approval of a large corporate loan requires a human loan officer to assess qualitative factors like leadership character. Similarly, in the legal field, core principles of justice and accountability demand that a human judge retain final authority over critical decisions like sentencing, which involve complex moral reasoning.
实施HITL模式使得智能体可以在完全自主不可行或不允许的敏感领域中使用。它还通过反馈循环提供了持续改进的机制。例如，在金融领域，大型企业贷款的最终批准需要人类贷款官员评估定性因素，如领导力特征。同样，在法律领域，正义和问责的核心原则要求人类法官保留对关键决策（如涉及复杂道德推理的量刑）的最终权威。
Caveats: Despite its benefits, the HITL pattern has significant caveats, chief among them being a lack of scalability. While human oversight provides high accuracy, operators cannot manage millions of tasks, creating a fundamental trade-off that often requires a hybrid approach combining automation for scale and HITL for accuracy. Furthermore, the effectiveness of this pattern is heavily dependent on the expertise of the human operators; for example, while an AI can generate software code, only a skilled developer can accurately identify subtle errors and provide the correct guidance to fix them. This need for expertise also applies when using HITL to generate training data, as human annotators may require special training to learn how to correct an AI in a way that produces high-quality data. Lastly, implementing HITL raises significant privacy concerns, as sensitive information must often be rigorously anonymized before it can be exposed to a human operator, adding another layer of process complexity.
注意事项： 尽管有好处，HITL模式也有显著的注意事项，其中最主要的是缺乏可扩展性。虽然人类监督提供了高准确性，但操作员无法管理数百万个任务，这造成了基本的权衡，通常需要结合自动化以实现规模化和HITL以实现准确性的混合方法。此外，这种模式的有效性在很大程度上取决于人类操作员的专业知识；例如，虽然人工智能可以生成软件代码，但只有熟练的开发人员才能准确识别细微错误并提供正确的修复指导。这种专业知识需求也适用于使用HITL生成训练数据时，因为人类标注员可能需要特殊培训来学习如何以产生高质量数据的方式纠正人工智能。最后，实施HITL会引发重大的隐私问题，因为敏感信息在暴露给人类操作员之前通常必须严格匿名化，这增加了另一层流程复杂性。
Practical Applications & Use Cases
实际应用与用例
The Human-in-the-Loop pattern is vital across a wide range of industries and applications, particularly where accuracy, safety, ethics, or nuanced understanding are paramount.
人在回路模式在广泛的行业和应用中至关重要，特别是在准确性、安全性、道德或细微理解至关重要的领域。
- Content Moderation: AI agents can rapidly filter vast amounts of online content for violations (e.g., hate speech, spam). However, ambiguous cases or borderline content are escalated to human moderators for review and final decision, ensuring nuanced judgment and adherence to complex policies.
- 内容审核： 人工智能智能体可以快速过滤大量在线内容以查找违规行为（例如，仇恨言论、垃圾邮件）。然而，模糊案例或边界内容会升级给人类审核员进行审查和最终决定，确保细微判断并遵守复杂政策。
- Autonomous Driving: While self-driving cars handle most driving tasks autonomously, they are designed to hand over control to a human driver in complex, unpredictable, or dangerous situations that the AI cannot confidently navigate (e.g., extreme weather, unusual road conditions).
- 自动驾驶： 虽然自动驾驶汽车自主处理大多数驾驶任务，但它们被设计为在人工智能无法自信导航的复杂、不可预测或危险情况下（例如，极端天气、异常路况）将控制权移交给人类驾驶员。
- Financial Fraud Detection: AI systems can flag suspicious transactions based on patterns. However, high-risk or ambiguous alerts are often sent to human analysts who investigate further, contact customers, and make the final determination on whether a transaction is fraudulent.
- 金融欺诈检测： 人工智能系统可以根据模式标记可疑交易。然而，高风险或模糊的警报通常会发送给人类分析师，他们进行进一步调查，联系客户，并最终确定交易是否具有欺诈性。
- Legal Document Review: AI can quickly scan and categorize thousands of legal documents to identify relevant clauses or evidence. Human legal professionals then review the AI's findings for accuracy, context, and legal implications, especially for critical cases.
- 法律文件审查： 人工智能可以快速扫描和分类数千份法律文件以识别相关条款或证据。然后，人类法律专业人员审查人工智能的发现，以确定准确性、上下文和法律含义，特别是对于关键案件。
- Customer Support (Complex Queries): A chatbot might handle routine customer inquiries. If the user's problem is too complex, emotionally charged, or requires empathy that the AI cannot provide, the conversation is seamlessly handed over to a human support agent.
- 客户支持（复杂查询）： 聊天机器人可能处理常规客户咨询。如果用户的问题过于复杂、情绪化或需要人工智能无法提供的同理心，对话会无缝地转交给人类支持代理。
- Data Labeling and Annotation: AI models often require large datasets of labeled data for training. Humans are put in the loop to accurately label images, text, or audio, providing the ground truth that the AI learns from. This is a continuous process as models evolve.
- 数据标注和注释： 人工智能模型通常需要大量带标签的数据集进行训练。人类被置于回路中，以准确标注图像、文本或音频，提供人工智能学习的基础事实。随着模型的发展，这是一个持续的过程。
- Generative AI Refinement: When an LLM generates creative content (e.g., marketing copy, design ideas), human editors or designers review and refine the output, ensuring it meets brand guidelines, resonates with the target audience, and maintains quality.
- 生成式人工智能精炼： 当大型语言模型生成创意内容（例如，营销文案、设计想法）时，人类编辑或设计师会审查和优化输出，确保其符合品牌指南，与目标受众产生共鸣，并保持质量。
- Autonomous Networks: AI systems are capable of analyzing alerts and forecasting network issues and traffic anomalies by leveraging key performance indicators (KPIs) and identified patterns. Nevertheless, crucial decisions—such as addressing high-risk alerts—are frequently escalated to human analysts. These analysts conduct further investigation and make the ultimate determination regarding the approval of network changes.
- 自主网络： 人工智能系统能够通过利用关键绩效指标（KPI）和已识别模式来分析警报并预测网络问题和流量异常。然而，关键决策——例如处理高风险警报——经常升级给人类分析师。这些分析师进行进一步调查，并就网络变更的批准做出最终决定。
This pattern exemplifies a practical method for AI implementation. It harnesses AI for enhanced scalability and efficiency, while maintaining human oversight to ensure quality, safety, and ethical compliance.
这种模式展示了人工智能实施的实用方法。它利用人工智能来增强可扩展性和效率，同时保持人类监督以确保质量、安全和道德合规性。
"Human-on-the-loop" is a variation of this pattern where human experts define the overarching policy, and the AI then handles immediate actions to ensure compliance. Let's consider two examples:
"人在环上"是这种模式的一个变体，其中人类专家定义总体政策，然后人工智能处理即时行动以确保合规性。让我们考虑两个例子：
- Automated financial trading system: In this scenario, a human financial expert sets the overarching investment strategy and rules. For instance, the human might define the policy as: "Maintain a portfolio of 70% tech stocks and 30% bonds, do not invest more than 5% in any single company, and automatically sell any stock that falls 10% below its purchase price." The AI then monitors the stock market in real-time, executing trades instantly when these predefined conditions are met. The AI is handling the immediate, high-speed actions based on the slower, more strategic policy set by the human operator.
- 自动化金融交易系统： 在这种情况下，人类金融专家设定总体投资策略和规则。例如，人类可能将政策定义为："保持投资组合为70%科技股和30%债券，不要在任何单一公司投资超过5%，并自动卖出任何低于购买价格10%的股票。"然后人工智能实时监控股票市场，在满足这些预定义条件时立即执行交易。人工智能基于人类操作员设定的较慢、更具战略性的政策来处理即时、高速的行动。
- Modern call center: In this setup, a human manager establishes high-level policies for customer interactions. For instance, the manager might set rules such as "any call mentioning 'service outage' should be immediately routed to a technical support specialist," or "if a customer's tone of voice indicates high frustration, the system should offer to connect them directly to a human agent." The AI system then handles the initial customer interactions, listening to and interpreting their needs in real-time. It autonomously executes the manager's policies by instantly routing the calls or offering escalations without needing human intervention for each individual case. This allows the AI to manage the high volume of immediate actions according to the slower, strategic guidance provided by the human operator.
- 现代呼叫中心： 在这种设置中，人类经理为客户互动建立高级政策。例如，经理可能设定规则，如"任何提到'服务中断'的电话应立即转接给技术支持专家"，或"如果客户的语气表明高度沮丧，系统应提供直接连接到人类代理的选项。"然后人工智能系统处理初始客户互动，实时听取和解释他们的需求。它通过即时转接电话或提供升级选项来自主执行经理的政策，无需为每个个案进行人工干预。这使得人工智能能够根据人类操作员提供的较慢、战略性指导来管理大量的即时行动。
Hands-On Code Example
实践代码示例
To demonstrate the Human-in-the-Loop pattern, an ADK agent can identify scenarios requiring human review and initiate an escalation process . This allows for human intervention in situations where the agent's autonomous decision-making capabilities are limited or when complex judgments are required. This is not an isolated feature; other popular frameworks have adopted similar capabilities. LangChain, for instance, also provides tools to implement these types of interactions.
为了演示人在回路模式，ADK智能体可以识别需要人工审查的场景并启动升级流程。这允许在智能体的自主决策能力有限或需要复杂判断的情况下进行人工干预。这不是一个孤立的功能；其他流行框架也采用了类似的能力。例如，LangChain也提供了实现这类互动的工具。
from google.adk.agents import Agent
from google.adk.tools.tool_context import ToolContext
from google.adk.callbacks import CallbackContext
from google.adk.models.llm import LlmRequest
from google.genai import types
from typing import Optional
# Placeholder for tools (replace with actual implementations if needed)
def troubleshoot_issue(issue: str) -> dict:
return {"status": "success", "report": f"Troubleshooting steps for {issue}."}
def create_ticket(issue_type: str, details: str) -> dict:
return {"status": "success", "ticket_id": "TICKET123"}
def escalate_to_human(issue_type: str) -> dict:
# This would typically transfer to a human queue in a real system
return {"status": "success", "message": f"Escalated {issue_type} to a human specialist."}
technical_support_agent = Agent(
name="technical_support_specialist",
model="gemini-2.0-flash-exp",
instruction="""
You are a technical support specialist for our electronics company.
FIRST, check if the user has a support history in state["customer_info"]["support_history"].
If they do, reference this history in your responses.
For technical issues:
1. Use the troubleshoot_issue tool to analyze the problem.
2. Guide the user through basic troubleshooting steps.
3. If the issue persists, use create_ticket to log the issue.
For complex issues beyond basic troubleshooting:
1. Use escalate_to_human to transfer to a human specialist.
Maintain a professional but empathetic tone. Acknowledge the frustration
technical issues can cause, while providing clear steps toward resolution.
""",
tools=[troubleshoot_issue, create_ticket, escalate_to_human]
)
def personalization_callback(
callback_context: CallbackContext,
llm_request: LlmRequest
) -> Optional[LlmRequest]:
"""Adds personalization information to the LLM request."""
# Get customer info from state
customer_info = callback_context.state.get("customer_info")
if customer_info:
customer_name = customer_info.get("name", "valued customer")
customer_tier = customer_info.get("tier", "standard")
recent_purchases = customer_info.get("recent_purchases", [])
personalization_note = (
f"\nIMPORTANT PERSONALIZATION:\n"
f"Customer Name: {customer_name}\n"
f"Customer Tier: {customer_tier}\n"
)
if recent_purchases:
personalization_note += f"Recent Purchases: {', '.join(recent_purchases)}\n"
if llm_request.contents:
# Add as a system message before the first content
system_content = types.Content(
role="system",
parts=[types.Part(text=personalization_note)]
)
llm_request.contents.insert(0, system_content)
return None # Return None to continue with the modified request
This code offers a blueprint for creating a technical support agent using Google's ADK, designed around a HITL framework. The agent acts as an intelligent first line of support, configured with specific instructions and equipped with tools like troubleshoot_issue, create_ticket, and escalate_to_human to manage a complete support workflow. The escalation tool is a core part of the HITL design, ensuring complex or sensitive cases are passed to human specialists.
这段代码提供了一个使用Google的ADK创建技术支持智能体的蓝图，围绕HITL框架设计。该智能体充当智能的第一线支持，配置了特定指令，并配备了troubleshoot_issue、create_ticket和escalate_to_human等工具来管理完整的支持工作流程。升级工具是HITL设计的核心部分，确保复杂或敏感案例传递给人类专家。
A key feature of this architecture is its capacity for deep personalization, achieved through a dedicated callback function. Before contacting the LLM, this function dynamically retrieves customer-specific data—such as their name, tier, and purchase history—from the agent's state. This context is then injected into the prompt as a system message, enabling the agent to provide highly tailored and informed responses that reference the user's history. By combining a structured workflow with essential human oversight and dynamic personalization, this code serves as a practical example of how the ADK facilitates the development of sophisticated and robust AI support solutions.
该架构的一个关键特性是其深度个性化能力，通过专用的回调函数实现。在联系大型语言模型之前，此函数从智能体状态中动态检索客户特定数据——例如他们的姓名、等级和购买历史。然后，此上下文作为系统消息注入提示中，使智能体能够提供高度定制和知情的响应，引用用户的历史记录。通过将结构化工作流程与基本的人类监督和动态个性化相结合，这段代码作为一个实际示例，展示了ADK如何促进复杂而强大的人工智能支持解决方案的开发。
At Glance
要点概览
What: AI systems, including advanced LLMs, often struggle with tasks that require nuanced judgment, ethical reasoning, or a deep understanding of complex, ambiguous contexts. Deploying fully autonomous AI in high-stakes environments carries significant risks, as errors can lead to severe safety, financial, or ethical consequences. These systems lack the inherent creativity and common-sense reasoning that humans possess. Consequently, relying solely on automation in critical decision-making processes is often imprudent and can undermine the system's overall effectiveness and trustworthiness.
是什么： 人工智能系统，包括先进的大型语言模型，通常难以处理需要细微判断、伦理推理或对复杂、模糊上下文有深刻理解的任务。在高风险环境中部署完全自主的人工智能具有重大风险，因为错误可能导致严重的安全、财务或道德后果。这些系统缺乏人类固有的创造力和常识推理能力。因此，在关键决策过程中仅依赖自动化通常是不明智的，可能损害系统的整体有效性和可信度。
Why: The Human-in-the-Loop (HITL) pattern provides a standardized solution by strategically integrating human oversight into AI workflows. This agentic approach creates a symbiotic partnership where AI handles computational heavy-lifting and data processing, while humans provide critical validation, feedback, and intervention. By doing so, HITL ensures that AI actions align with human values and safety protocols. This collaborative framework not only mitigates the risks of full automation but also enhances the system's capabilities through continuous learning from human input. Ultimately, this leads to more robust, accurate, and ethical outcomes that neither human nor AI could achieve alone.
为什么： 人在回路（HITL）模式通过战略性地将人类监督整合到人工智能工作流程中，提供了一个标准化解决方案。这种智能体方法创建了一种共生伙伴关系，其中人工智能处理计算密集型工作和数据处理，而人类提供关键验证、反馈和干预。通过这样做，HITL确保人工智能行动与人类价值观和安全协议保持一致。这种协作框架不仅减轻了完全自动化的风险，还通过从人类输入中持续学习来增强系统能力。最终，这导致了更强大、准确和道德的结果，这是人类或人工智能单独都无法实现的。
Rule of thumb: Use this pattern when deploying AI in domains where errors have significant safety, ethical, or financial consequences, such as in healthcare, finance, or autonomous systems. It is essential for tasks involving ambiguity and nuance that LLMs cannot reliably handle, like content moderation or complex customer support escalations. Employ HITL when the goal is to continuously improve an AI model with high-quality, human-labeled data or to refine generative AI outputs to meet specific quality standards.
经验法则： 在错误具有重大安全、道德或财务后果的领域（如医疗保健、金融或自主系统）部署人工智能时使用此模式。对于涉及大型语言模型无法可靠处理的模糊性和细微差别的任务（如内容审核或复杂客户支持升级）至关重要。当目标是通过高质量、人类标注的数据持续改进人工智能模型，或精炼生成式人工智能输出以满足特定质量标准时，采用HITL。
Visual summary:
Fig.1: Human in the loop design pattern 图1：人在回路设计模式
Key Takeaways
关键结论
Key takeaways include:
关键结论包括：
- Human-in-the-Loop (HITL) integrates human intelligence and judgment into AI workflows.
- 人在回路（HITL） 将人类智能和判断整合到人工智能工作流程中。
- It's crucial for safety, ethics, and effectiveness in complex or high-stakes scenarios.
- 在复杂或高风险场景中，它对安全、道德和有效性至关重要。
- Key aspects include human oversight, intervention, feedback for learning, and decision augmentation.
- 关键方面包括人类监督、干预、用于学习的反馈和决策增强。
- Escalation policies are essential for agents to know when to hand off to a human.
- 升级策略对于智能体知道何时转交给人类至关重要。
- HITL allows for responsible AI deployment and continuous improvement.
- HITL允许负责任的人工智能部署和持续改进。
- The primary drawbacks of Human-in-the-Loop are its inherent lack of scalability, creating a trade-off between accuracy and volume, and its dependence on highly skilled domain experts for effective intervention.
- 人在回路的主要缺点是其固有的缺乏可扩展性，造成准确性和数量之间的权衡，以及其依赖高技能领域专家进行有效干预。
- Its implementation presents operational challenges, including the need to train human operators for data generation and to address privacy concerns by anonymizing sensitive information.
- 其实施带来了操作挑战，包括需要培训人类操作员进行数据生成，以及通过匿名化敏感信息来解决隐私问题。
Conclusion
结论
This chapter explored the vital Human-in-the-Loop (HITL) pattern, emphasizing its role in creating robust, safe, and ethical AI systems. We discussed how integrating human oversight, intervention, and feedback into agent workflows can significantly enhance their performance and trustworthiness, especially in complex and sensitive domains. The practical applications demonstrated HITL's widespread utility, from content moderation and medical diagnosis to autonomous driving and customer support. The conceptual code example provided a glimpse into how ADK can facilitate these human-agent interactions through escalation mechanisms. As AI capabilities continue to advance, HITL remains a cornerstone for responsible AI development, ensuring that human values and expertise remain central to intelligent system design.
本章探讨了至关重要的人在回路（HITL）模式，强调了其在创建强大、安全和道德的人工智能系统中的作用。我们讨论了如何将人类监督、干预和反馈整合到智能体工作流程中可以显著提高其性能和可信度，特别是在复杂和敏感的领域。实际应用展示了HITL的广泛实用性，从内容审核和医疗诊断到自动驾驶和客户支持。概念代码示例提供了ADK如何通过升级机制促进这些人机互动的一瞥。随着人工智能能力的不断进步，HITL仍然是负责任人工智能开发的基石，确保人类价值观和专业知识在智能系统设计中保持核心地位。
References
参考文献
- A Survey of Human-in-the-loop for Machine Learning, Xingjiao Wu, Luwei Xiao, Yixuan Sun, Junhang Zhang, Tianlong Ma, Liang He, arxiv.org/abs/2108.00…
- 人在回路机器学习综述，吴兴蛟，肖路伟，孙逸轩，张俊航，马天龙，何亮，arxiv.org/abs/2108.00…
