麦肯锡-为什么代理是生成式人工智能的下一个前沿（英）-2024-WN8.doc

McKinseyDigitalPractice

Whyagentsarethenext

frontierofgenerativeAI

Bymovingfrominformationtoaction—thinkvirtualcoworkers

abletocompletecomplexworkflows—thetechnologypromisesa

newwaveofproductivityandinnovation.

ByLareinaYee,MichaelChui,andRogerRoberts

withStephenXu

July2024

Overthepastcoupleofyears,theworldhas

marveledatthecapabilitiesandpossibilities

unleashedbygenerativeAI(genAI).Foundation

modelssuchaslargelanguagemodels(LLMs)can

performimpressivefeats,extractinginsightsand

generatingcontentacrossnumerousmediums,

suchastext,audio,images,andvideo.Butthenext

stageofgenAIislikelytobemoretransformative.

agentsystemtoaccomplishacomplexwork?ow.

Amultiagentsystemcouldtheninterpretand

organizethiswork?owintoactionabletasks,assign

worktospecializedagents,executethesere?ned

tasksusingadigitalecosystemoftools,and

collaboratewithotheragentsandhumansto

iterativelyimprovethequalityofitsactions.

Wearebeginninganevolutionfrom

knowledge-based,genAI–powered

tools—say,chatbotsthatanswerquestions

andgeneratecontent—togenAI–enabled“agents”

thatusefoundationmodelstoexecutecomplex,

multistepwork?owsacrossadigitalworld.Inshort,

thetechnologyismovingfromthoughttoaction.

Broadlyspeaking,“agentic”systemsreferto

digitalsystemsthatcanindependentlyinteract

inadynamicworld.Whileversionsofthese

softwaresystemshaveexistedforyears,the

natural-languagecapabilitiesofgenAIunveilnew

possibilities,enablingsystemsthatcanplantheir

actions,useonlinetoolstocompletethosetasks,

collaboratewithotheragentsandpeople,andlearn

toimprovetheirperformance.GenAIagents

eventuallycouldactasskilledvirtualcoworkers,

workingwithhumansinaseamlessandnatural

manner.Avirtualassistant,forexample,couldplan

andbookacomplexpersonalizedtravelitinerary,

han