多智能体AI——21世纪的自动化革命

Multi-agent AI– 21stcenturyautomation revolution Automation has been the fundamental technologyunderpinning economic transformation for centuries. Whether it was Britain’s late 18thand 19thcentury industrial revolution, the UnitedStates’ post-World War II boom, or SouthKorea’s industrialization starting in the1960s, all benefited from automationto increase productivity, efficiency,and profitability. The effect was radicaltransformation of economy and society.1Now, in the 21stcentury, a new wave ofautomation is underway as agentic AI usewidens across economies. AI agents employ a range of advancedtechnologies to interact with usersand perform tasks autonomously andeffectively. Large language models (LLMs)are often the primary interface betweenAI agents and users. They are a type offoundation model trained on vast datasets,primarily text. LLMs encode knowledge byrecognizing patterns, enabling AI agentsto reason, inform their decision-making,and communicate. Anartificial intelligence agentis a softwareprogram that can interact with itsenvironment, collect data, and use thisto autonomously perform tasks to meetpredetermined goals. As an evolutionfrom technologies like robotic processautomation (RPA) and machine learning(ML), AI agents can, perceive, reason, andact in changing environments to achievetheir goals. How they reach them is largelyleft to them to decide. An agent can understand and generatehuman-like text or verbal responsesusing natural language processing(NLP), making human-AI interactionsmore natural and efficient. Agentic AIdoing what it does best Since generative AI took off as a popular phenomenon,companies have rushed to create their own versions.They are used for advanced search, analysis andinteraction with documents, particularly in the legal,HR, and technology fields. While these improve onprevious systems, to stop at this point underestimatesthe full potential. LLMs are improving iteratively –where two years ago an LLM with retrieval-augmentedgeneration (RAG) could produce synopses – currentLLMs using enhanced retrieval methods can generatemore sophisticated output. The fundamental difference between using a standaloneLLM and employing a multi-agent system is that withthe latter, individual agents are created to specialize inspecific tasks – often not limited to language – and cancollaborate with each other. They can execute morecomplex tasks and integrate with external tools suchas web searches, APIs, and dedicated databases. Tangible and intangiblevalue Why should enterprises go the agentic AI route? In short,because it will soon be in every business function whereit’s feasible. Companies are re-organizing their customer and ITsupport services to raise quality standards by augmentingthem with AI agents in hybrid models. For example, anAI agent can automatically draft responses to customerqueries based on historical customer interaction data. 82%of organizationsplan to integrateAI agents by 2027.2 More ambitiously, agents can take ownership of a clientissue. Previously, if a customer contacted a chatbot torequest a refund for a product and the request wasnon-standard, the chatbot would escalate the issue toa human customer-service representative. Now, an AIagent can request more information from a customer,for example, proof of purchase or a photograph of anitem, analyze the enquiry and offer a possible solution.It can decide alone to override standard procedure, ifcircumstances justify making an exception. It is highlylikely to resolve the matter without human input. There are more reasons to move to agentic AI thancost-saving alone. According to a UK Customer Satisfaction survey, adefining feature of high-performing organizations is“anappropriate balance of people and technology, combiningspeed, efficiency, and personal care.” In other words, the ideal retail customer serviceexperience is a blend of human empathy andtechnological efficiency.3Businesses can thereforeuse agentic AI to improve and differentiate their offerto customers ahead of competitors, e.g., by addingcommunication channels and styles that appeal tospecific customer bases. Design formaximum return An AI agentic system increases overall productivity, service quality, customersatisfaction, and loyalty. The design of agent-to-agent and agent-to-humanexchanges according to frameworks, rules, risks, and protocols is crucial for this. Unlike predecessors, AI agents are: •Autonomous•Goal-oriented•Context-aware, using relevant data to make decisions•Adaptive, adjusting behavior and responses as data or interactions change•Proactive, initiating action independently without user prompts•Language-aware, interpreting and responding in human language Agentcreation To build AI agents requires: Typical goals for agents: 1.Defining their roles,2.Identifying and locating the data they will use,3.Defining which tasks or goals they execute,4.Setting boundaries with guardrails. Contentrecommendation Customerservic