AI智能体：概念、运作原理及组织最优应用场景

AI Agents: What They Are,How They Work, and WhereOrganizations Might Best Use Them By: Matt Mullen based on complexity. Finally, we discuss thespecific use cases that an enterprise mightconsider appropriate for deploying agents andhow the preparation, implementation, andoperation of such use might look. The Essentials In 2024, the introduction of “agents” derivedfrom artificial intelligence (AI) dominated therelatively immature marketplace in generativeAI for enterprises. Whereas in 2023, discussionof enterprise AI centered upon its use as anassistantwithin business applications, 2024saw automation, process orchestration, andintegration built onto that foundation, creatingthe notion ofagents. Agents can enable enterprises to shift intoa level of complex process orchestrationthat previously would have been resourceprohibitive (in both cost and skill). To takeadvantage of these capabilities, a priorityplanning task is creating and supporting theactions needed to make plans executable byprospective agents. The specificity that makesagents useful means that every action they areapproved to perform needs to be supported tosome extent by additional enterprise software(for example, business applications) andcustomer-specific configuration. Using increasingly sophisticated, specialistmodels and enterprise data along withsoftware engines that can make determinationson purpose, plan, and execution, agentspromise a high degree ofspecificityin carryingout sets of tasks. Where previous iterationsof technology had to rely on highly detailedpredetermined plans, agents can create theirown plans based on sets of actions that theyhave been cleared to operate. Traversing the agent landscape as outlinedwithin this report is a vital first step fororganizations that recognize their ownoperational challenges among those proposedby the industry. Successful execution willbring a bounty of potential benefits, providedthat execution is based on sound planningand successful harnessing of human skillsand experience. In this report, we examine the characteristicsof agents and how they extend the capabilitiesof assistants to achieve more specific, complexoutcomes. We also look at the operational andtechnical architectures that support agents’operation and classify the supported actions Finally, looking at the short to medium term –the next 12-24 months – we acknowledge thatinitially, agents will be deployed primarily byorganizations with existing, application-specificworkflows and codebases that can be utilizedright away within single-vendor environments.From this starting point, the provisioning oflibraries of actions from which agents can beassembled and configured will become partof vendor-provided capabilities, with gradualintegration support for third-party applicationsemerging. This might seem an unambitioustrajectory, but it will still be challenging for evenhighly engaged organizations to achieve. Figure 1 Background:From Assistants to Agents In October 2023, Deep Analysis releasedthe report “Workplace AI Market Analysis:Generative AI and the Desktop (R)Evolution”,1which detailed how relatively new technologicaladvances had enabled generative AI (GenAI)assistants to become a brand-new element inmany desktop software applications. A yearlater, at the time of writing this report, manyof those software applications are alreadyaugmenting those existing – still fresh –assistants with agents, often with very similardescriptions and overlapping messaging as totheir capabilities. in doing so, significant existing elements of anorganization’s information architecture startdrawing much closer together. As we will discuss, agents enableInformation,Transaction, and Extendeduse cases, whichas they increase in complexity also deliver anequivalent amount of specificity. The language of agents In part because agents came into being asproducts so hot on the heels of AI assistants,it has been difficult for many to comfortablyarticulate ways to distinguish between the two.Some software vendors have iterated theirproduct naming and descriptions to fold agentcapabilities into existing branding, while othershave thrown out their branding, introducedsomething new, and renamed previouslylaunched products accordingly. Confusion wasperhaps not the goal, yet it has certainly beenan unintended consequence. That report attempted to explain the movingparts around assistants – including the mostcited use cases and suggested approaches forimplementation – and this report attempts todo the same for this next market iteration. AIagents represent a significant advance fromthe scope largely attributed to assistants, andthis report seeks to explain that advance, howagents extend the vision for assistance muchfurther than the “assistant” iteration, and how For clarity, this report will primarily refer toagents,reasoning engines,andactions(seeFigure 1). Figure 2 Agentsare software entities that interactwith human or machine requesters tocollect, re