2025年Agents與基礎(chǔ)應(yīng)用白皮書（英文版）-谷歌

Agents

Authors:JuliaWiesinger,PatrickMarlowandVladimirVuskovic

Google

Agents

Acknowledgements

ReviewersandContributors

EvanHuangEmilyXue

OlcanSercinogluSebastianRiedelSatinderBaveja

AntonioGulli

AnantNawalgaria

CuratorsandEditors

AntonioGulli

AnantNawalgariaGraceMollison

TechnicalWriter

JoeyHaymaker

Designer

MichaelLanning

September20242

Tableofcontents

Introduction4

Whatisanagent?5

Themodel·6

Thetools·7

Theorchestrationlayer7

Agentsvs.models·8

Cognitivearchitectures:Howagentsoperate··8

Tools:Ourkeystotheoutsideworld12

Extensions··13

SampleExtensions··15

Functions·18

Usecases21

Functionsamplecode24

Datastores·27

Implementationandapplication28

Toolsrecap··32

Enhancingmodelperformancewithtargetedlearning·33

AgentquickstartwithLangChain35

ProductionapplicationswithVertexAIagents38

Summary40

Endnotes42

Agents

Thiscombinationofreasoning,

logic,andaccesstoexternal

informationthatareallconnectedtoaGenerativeAImodelinvokestheconceptofanagent.

Introduction

Humansarefantasticatmessypatternrecognitiontasks.However,theyoftenrelyontools-likebooks,GoogleSearch,oracalculator-tosupplementtheirpriorknowledgebefore

arrivingataconclusion.Justlikehumans,GenerativeAImodelscanbetrainedtousetoolstoaccessreal-timeinformationorsuggestareal-worldaction.Forexample,amodelcan

leverageadatabaseretrievaltooltoaccessspecificinformation,likeacustomer'spurchasehistory,soitcangeneratetailoredshoppingrecommendations.Alternatively,basedona

user'squery,amodelcanmakevariousAPIcallstosendanemailresponsetoacolleagueorcompleteafinancialtransactiononyourbehalf.Todoso,themodelmustnotonlyhaveaccesstoasetofexternaltools,itneedstheabilitytoplanandexecuteanytaskinaself-directedfashion.Thiscombinationofreasoning,logic,andaccesstoexternalinformation

thatareallconnectedtoaGenerativeAImodelinvokestheconceptofanagent,ora

programthatextendsbeyondthestandalonecapabilitiesofaGenerativeAImodel.Thiswhitepaperdivesintoalltheseandassociatedaspectsinmoredetail.

September20244

Agents

Whatisanagent?

Initsmostfundamentalform,aGenerativeAIagentcanbedefinedasanapplicationthat

attemptstoachieveagoalbyobservingtheworldandactinguponitusingthetoolsthatit

hasatitsdisposal.Agentsareautonomousandcanactindependentlyofhumanintervention,especiallywhenprovidedwithpropergoalsorobjectivestheyaremeanttoachieve.Agentscanalsobeproactiveintheirapproachtoreachingtheirgoals.Evenintheabsenceof

explicitinstructionsetsfromahuman,anagentcanreasonaboutwhatitshoulddonexttoachieveitsultimategoal.WhilethenotionofagentsinAIisquitegeneralandpowerful,thiswhitepaperfocusesonthespecifictypesofagentsthatGenerativeAImodelsarecapableofbuildingatthetimeofpublication.

Inordertounderstandtheinnerworkingsofanagent,let’sfirstintroducethefoundationalcomponentsthatdrivetheagent’sbehavior,actions,anddecisionmaking.Thecombinationofthesecomponentscanbedescribedasacognitivearchitecture,andtherearemany

sucharchitecturesthatcanbeachievedbythemixingandmatchingofthesecomponents.Focusingonthecorefunctionalities,therearethreeessentialcomponentsinanagent’s

cognitivearchitectureasshowninFigure1.

September2024s

Agents

Figure1.Generalagentarchitectureandcomponents

Themodel

Inthescopeofanagent,amodelreferstothelanguagemodel(LM)thatwillbeutilizedas

thecentralizeddecisionmakerforagentprocesses.ThemodelusedbyanagentcanbeoneormultipleLM’sofanysize(small/large)thatarecapableoffollowinginstructionbased

reasoningandlogicframeworks,likeReAct,Chain-of-Thought,orTree-of-Thoughts.Modelscanbegeneralpurpose,multimodalorfine-tunedbasedontheneedsofyourspecificagentarchitecture.Forbestproductionresults,youshouldleverageamodelthatbestfitsyour

desiredendapplicationand,ideally,hasbeentrainedondatasignaturesassociatedwiththetoolsthatyouplantouseinthecognitivearchitecture.It’simportanttonotethatthemodelistypicallynottrainedwiththespecificconfigurationsettings(i.e.toolchoices,orchestration/reasoningsetup)oftheagent.However,it’spossibletofurtherrefinethemodelforthe

agent’stasksbyprovidingitwithexamplesthatshowcasetheagent’scapabilities,includinginstancesoftheagentusingspecifictoolsorreasoningstepsinvariouscontexts.

September20246

Agents

Thetools

Foundationalmodels,despitetheirimpressivetextandimagegeneration,remainconstrainedbytheirinabilitytointeractwiththeoutsideworld.Toolsbridgethisgap,empoweringagentstointeractwithexternaldataandserviceswhileunlockingawiderrangeofactionsbeyond

thatoftheunderlyingmodelalone.Toolscantakeavarietyofformsandhavevarying

depthsofcomplexity,buttypicallyalignwithcommonwebAPImethodslikeGET,POST,

PATCH,andDELETE.Forexample,atoolcouldupdatecustomerinformationinadatabaseorfetchweatherdatatoinfluenceatravelrecommendationthattheagentisprovidingtotheuser.Withtools,agentscanaccessandprocessreal-worldinformation.Thisempowers

themtosupportmorespecializedsystemslikeretrievalaugmentedgeneration(RAG),

whichsignificantlyextendsanagent’scapabilitiesbeyondwhatthefoundationalmodelcanachieveonitsown.We’lldiscusstoolsinmoredetailbelow,butthemostimportantthingtounderstandisthattoolsbridgethegapbetweentheagent’sinternalcapabilitiesandtheexternalworld,unlockingabroaderrangeofpossibilities.

Theorchestrationlayer

Theorchestrationlayerdescribesacyclicalprocessthatgovernshowtheagenttakesin

information,performssomeinternalreasoning,andusesthatreasoningtoinformitsnextactionordecision.Ingeneral,thisloopwillcontinueuntilanagenthasreacheditsgoalorastoppingpoint.Thecomplexityoftheorchestrationlayercanvarygreatlydependingontheagentandtaskit’sperforming.Someloopscanbesimplecalculationswithdecisionrules,whileothersmaycontainchainedlogic,involveadditionalmachinelearningalgorithms,orimplementotherprobabilisticreasoningtechniques.We’lldiscussmoreaboutthedetailedimplementationoftheagentorchestrationlayersinthecognitivearchitecturesection.

September20247

Agents

Agentsvs.models

Togainaclearerunderstandingofthedistinctionbetweenagentsandmodels,considerthefollowingchart:

Models

Agents

Knowledgeislimitedtowhatisavailableintheirtrainingdata.

Knowledgeisextendedthroughtheconnectionwithexternalsystemsviatools

Singleinference/predictionbasedonthe

userquery.Unlessexplicitlyimplementedforthemodel,thereisnomanagementofsessionhistoryorcontinuouscontext.(i.e.chathistory)

Managedsessionhistory(i.e.chathistory)to

allowformultiturninference/predictionbasedonuserqueriesanddecisionsmadeinthe

orchestrationlayer.Inthiscontext,a‘turn’is

definedasaninteractionbetweentheinteractingsystemandtheagent.(i.e.1incomingevent/

queryand1agentresponse)

Nonativetoolimplementation.

Toolsarenativelyimplementedinagentarchitecture.

Nonativelogiclayerimplemented.Userscanformpromptsassimplequestionsoruse

reasoningframeworks(CoT,ReAct,etc.)toformcomplexpromptstoguidethemodelin

prediction.

NativecognitivearchitecturethatusesreasoningframeworkslikeCoT,ReAct,orotherpre-built

agentframeworkslikeLangChain.

Cognitivearchitectures:Howagentsoperate

Imagineachefinabusykitchen.Theirgoalistocreatedeliciousdishesforrestaurantpatronswhichinvolvessomecycleofplanning,execution,andadjustment.

September20248

Agents

?Theygatherinformation,likethepatron’sorderandwhatingredientsareinthepantryandrefrigerator.

?Theyperformsomeinternalreasoningaboutwhatdishesandflavorprofilestheycancreatebasedontheinformationtheyhavejustgathered.

?Theytakeactiontocreatethedish:choppingvegetables,blendingspices,searingmeat.

Ateachstageintheprocessthechefmakesadjustmentsasneeded,refiningtheirplanasingredientsaredepletedorcustomerfeedbackisreceived,andusesthesetofprevious

outcomestodeterminethenextplanofaction.Thiscycleofinformationintake,planning,executing,andadjustingdescribesauniquecognitivearchitecturethatthechefemploystoreachtheirgoal.

Justlikethechef,agentscanusecognitivearchitecturestoreachtheirendgoalsby

iterativelyprocessinginformation,makinginformeddecisions,andrefiningnextactions

basedonpreviousoutputs.Atthecoreofagentcognitivearchitecturesliestheorchestrationlayer,responsibleformaintainingmemory,state,reasoningandplanning.Itusestherapidlyevolvingfieldofpromptengineeringandassociatedframeworkstoguidereasoningand

planning,enablingtheagenttointeractmoreeffectivelywithitsenvironmentandcompletetasks.Researchintheareaofpromptengineeringframeworksandtaskplanningfor

languagemodelsisrapidlyevolving,yieldingavarietyofpromisingapproaches.Whilenotanexhaustivelist,theseareafewofthemostpopularframeworksandreasoningtechniquesavailableatthetimeofthispublication:

?ReAct,apromptengineeringframeworkthatprovidesathoughtprocessstrategyfor

languagemodelstoReasonandtakeactiononauserquery,withorwithoutin-context

examples.ReActpromptinghasshowntooutperformseveralSOTAbaselinesandimprovehumaninteroperabilityandtrustworthinessofLLMs.

September20249

Agents

?Chain-of-Thought(CoT),apromptengineeringframeworkthatenablesreasoning

capabilitiesthroughintermediatesteps.Therearevarioussub-techniquesofCoTincludingself-consistency,active-prompt,andmultimodalCoTthateachhavestrengthsand

weaknessesdependingonthespecificapplication.

?Tree-of-thoughts(ToT),,apromptengineeringframeworkthatiswellsuitedfor

explorationorstrategiclookaheadtasks.Itgeneralizesoverchain-of-thoughtpromptingandallowsthemodeltoexplorevariousthoughtchainsthatserveasintermediatestepsforgeneralproblemsolvingwithlanguagemodels.

Agentscanutilizeoneoftheabovereasoningtechniques,ormanyothertechniques,to

choosethenextbestactionforthegivenuserrequest.Forexample,let’sconsideranagentthatisprogrammedtousetheReActframeworktochoosethecorrectactionsandtoolsfortheuserquery.Thesequenceofeventsmightgosomethinglikethis:

1.Usersendsquerytotheagent

2.AgentbeginstheReActsequence

3.Theagentprovidesaprompttothemodel,askingittogenerateoneofthenextReActstepsanditscorrespondingoutput:

a.Question:Theinputquestionfromtheuserquery,providedwiththeprompt

b.Thought:Themodel’sthoughtsaboutwhatitshoulddonext

c.Action:Themodel’sdecisiononwhatactiontotakenexti.Thisiswheretoolchoicecanoccur

ii.Forexample,anactioncouldbeoneof[Flights,Search,Code,None],wherethefirst3representaknowntoolthatthemodelcanchoose,andthelastrepresents“no

toolchoice”

September202410

Agents

d.Actioninput:Themodel’sdecisiononwhatinputstoprovidetothetool(ifany)e.Observation:Theresultoftheaction/actioninputsequence

i.Thisthought/action/actioninput/observationcouldrepeatN-timesasneededf.Finalanswer:Themodel’sfinalanswertoprovidetotheoriginaluserquery

4.TheReActloopconcludesandafinalanswerisprovidedbacktotheuser

Figure2.ExampleagentwithReActreasoningintheorchestrationlayer

AsshowninFigure2,themodel,tools,andagentconfigurationworktogethertoprovideagrounded,conciseresponsebacktotheuserbasedontheuser’soriginalquery.Whilethemodelcouldhaveguessedatananswer(hallucinated)basedonitspriorknowledge,itinsteadusedatool(Flights)tosearchforreal-timeexternalinformation.Thisadditional

informationwasprovidedtothemodel,allowingittomakeamoreinformeddecisionbasedonrealfactualdataandtosummarizethisinformationbacktotheuser.

September202411

Agents

Insummary,thequalityofagentresponsescanbetieddirectlytothemodel’sabilityto

reasonandactaboutthesevarioustasks,includingtheabilitytoselecttherighttools,andhowwellthattoolshasbeendefined.Likeachefcraftingadishwithfreshingredientsandattentivetocustomerfeedback,agentsrelyonsoundreasoningandreliableinformationtodeliveroptimalresults.Inthenextsection,we’lldiveintothevariouswaysagentsconnectwithfreshdata.

Tools:Ourkeystotheoutsideworld

Whilelanguagemodelsexcelatprocessinginformation,theylacktheabilitytodirectly

perceiveandinfluencetherealworld.Thislimitstheirusefulnessinsituationsrequiring

interactionwithexternalsystemsordata.Thismeansthat,inasense,alanguagemodel

isonlyasgoodaswhatithaslearnedfromitstrainingdata.Butregardlessofhowmuch

datawethrowatamodel,theystilllackthefundamentalabilitytointeractwiththeoutsideworld.Sohowcanweempowerourmodelstohavereal-time,context-awareinteractionwithexternalsystems?Functions,Extensions,DataStoresandPluginsareallwaystoprovidethiscriticalcapabilitytothemodel.

Whiletheygobymanynames,toolsarewhatcreatealinkbetweenourfoundationalmodelsandtheoutsideworld.Thislinktoexternalsystemsanddataallowsouragenttoperformawidervarietyoftasksanddosowithmoreaccuracyandreliability.Forinstance,toolscanenableagentstoadjustsmarthomesettings,updatecalendars,fetchuserinformationfromadatabase,orsendemailsbasedonaspecificsetofinstructions.

Asofthedateofthispublication,therearethreeprimarytooltypesthatGooglemodelsareabletointeractwith:Extensions,Functions,andDataStores.Byequippingagentswithtools,weunlockavastpotentialforthemtonotonlyunderstandtheworldbutalsoactuponit,

openingdoorstoamyriadofnewapplicationsandpossibilities.

September202412

Agents

Extensions

TheeasiestwaytounderstandExtensionsistothinkofthemasbridgingthegapbetween

anAPIandanagentinastandardizedway,allowingagentstoseamlesslyexecuteAPIs

regardlessoftheirunderlyingimplementation.Let’ssaythatyou’vebuiltanagentwithagoalofhelpingusersbookflights.YouknowthatyouwanttousetheGoogleFlightsAPItoretrieveflightinformation,butyou’renotsurehowyou’regoingtogetyouragenttomakecallstothisAPIendpoint.

Figure3.HowdoAgentsinteractwithExternalAPIs?

Oneapproachcouldbetoimplementcustomcodethatwouldtaketheincominguserquery,parsethequeryforrelevantinformation,thenmaketheAPIcall.Forexample,inaflight

bookingusecaseausermightstate“IwanttobookaflightfromAustintoZurich.”Inthis

scenario,ourcustomcodesolutionwouldneedtoextract“Austin”and“Zurich”asrelevantentitiesfromtheuserquerybeforeattemptingtomaketheAPIcall.Butwhathappensiftheusersays“IwanttobookaflighttoZurich”andneverprovidesadeparturecity?TheAPIcall

wouldfailwithouttherequireddataandmorecodewouldneedtobeimplementedinordertocatchedgeandcornercaseslikethis.Thisapproachisnotscalableandcouldeasilybreakinanyscenariothatfallsoutsideoftheimplementedcustomcode.

September202413

Agents

AmoreresilientapproachwouldbetouseanExtension.AnExtensionbridgesthegapbetweenanagentandanAPIby:

1.TeachingtheagenthowtousetheAPIendpointusingexamples.

2.TeachingtheagentwhatargumentsorparametersareneededtosuccessfullycalltheAPIendpoint.

Figure4.ExtensionsconnectAgentstoExternalAPIs

Extensionscanbecraftedindependentlyoftheagent,butshouldbeprovidedaspartoftheagent’sconfiguration.TheagentusesthemodelandexamplesatruntimetodecidewhichExtension,ifany,wouldbesuitableforsolvingtheuser’squery.ThishighlightsakeystrengthofExtensions,theirbuilt-inexampletypes,thatallowtheagenttodynamicallyselectthe

mostappropriateExtensionforthetask.

Figure5.1-to-manyrelationshipbetweenAgents,ExtensionsandAPIs

September202414

Agents

ThinkofthisthesamewaythatasoftwaredeveloperdecideswhichAPIendpointstousewhilesolvingandsolutioningforauser’sproblem.Iftheuserwantstobookaflight,the

developermightusetheGoogleFlightsAPI.Iftheuserwantstoknowwherethenearest

coffeeshopisrelativetotheirlocation,thedevelopermightusetheGoogleMapsAPI.In

thissameway,theagent/modelstackusesasetofknownExtensionstodecidewhichonewillbethebestfitfortheuser’squery.Ifyou’dliketoseeExtensionsinaction,youcantrythemoutontheGeminiapplicationbygoingtoSettings>Extensionsandthenenablinganyyouwouldliketotest.Forexample,youcouldenabletheGoogleFlightsextensionthenaskGemini“ShowmeflightsfromAustintoZurichleavingnextFriday.”

SampleExtensions

TosimplifytheusageofExtensions,Googleprovidessomeoutoftheboxextensionsthat

canbequicklyimportedintoyourprojectandusedwithminimalconfigurations.Forexample,theCodeInterpreterextensioninSnippet1allowsyoutogenerateandrunPythoncodefromanaturallanguagedescription.

September202415

Agents

Python

importvertexaiimportpprint

PROJECT_ID="YOUR_PROJECT_ID"REGION="us-central1"

vertexai.init(project=PROJECT_ID,location=REGION)

fromvertexai.preview.extensionsimportExtension

extension_code_interpreter=Extension.from_hub("code_interpreter")

CODE_QUERY="""WriteapythonmethodtoinvertabinarytreeinO(n)time."""

response=extension_code_interpreter.execute(

operation_id="generate_and_execute",

operation_params={"query":CODE_QUERY})

print("GeneratedCode:")

pprint.pprint({response['generated_code']})

#Theabovesnippetwillgeneratethefollowingcode.

```

GeneratedCode:classTreeNode:

definit(self,val=0,left=None,right=None):self.val=val

self.left=left

self.right=right

Continuesnextpage...

September202416

Agents

Python

definvert_binary_tree(root):

"""

Invertsabinarytree.Args:

root:Therootofthebinarytree.Returns:

Therootoftheinvertedbinarytree.

"""

ifnotroot:

returnNone

#Swaptheleftandrightchildrenrecursively

root.left,root.right=

invert_binary_tree(root.right),invert_binary_tree(root.left)

returnroot

#Exampleusage:

#Constructasamplebinarytree

root=TreeNode(4)

root.left=TreeNode(2)root.right=TreeNode(7)

root.left.left=TreeNode(1)

root.left.right=TreeNode(3)root.right.left=TreeNode(6)root.right.right=TreeNode(9)

#Invertthebinarytree

inverted_root=invert_binary_tree(root)```

Snippet1.CodeInterpreterExtensioncangenerateandrunPythoncode

September202417

Agents

Tosummarize,Extensionsprovideawayforagentstoperceive,interact,andinfluencetheoutsideworldinamyriadofways.TheselectionandinvocationoftheseExtensionsisguidedbytheuseofExamples,allofwhicharedefinedaspartoftheExtensionconfiguration.

Functions

Intheworldofsoftwareengineering,functionsaredefinedasself-containedmodules

ofcodethataccomplishaspecifictaskandcanbereusedasneeded.Whenasoftwaredeveloperiswritingaprogram,theywilloftencreatemanyfunctionstodovarioustasks.Theywillalsodefinethelogicforwhentocallfunction_aversusfunction_b,aswellastheexpectedinputsandoutputs.

Functionsworkverysimilarlyintheworldofagents,butwecanreplacethesoftware

developerwithamodel.AmodelcantakeasetofknownfunctionsanddecidewhentouseeachFunctionandwhatargumentstheFunctionneedsbasedonitsspecification.FunctionsdifferfromExtensionsinafewways,mostnotably:

1.AmodeloutputsaFunctionanditsarguments,butdoesn’tmakealiveAPIcall.

2.Functionsareexecutedontheclient-side,whileExtensionsareexecutedontheagent-side.

UsingourGoogleFlightsexampleagain,asimplesetupforfunctionsmightlookliketheexampleinFigure7.

September202418

Agents

Figure7.HowdofunctionsinteractwithexternalAPIs?

NotethatthemaindifferencehereisthatneithertheFunctionnortheagentinteractdirectlywiththeGoogleFlightsAPI.SohowdoestheAPIcallactuallyhappen?

Withfunctions,thelogicandexecutionofcallingtheactualAPIendpointisoffloadedawayfromtheagentandbacktotheclient-sideapplicationasseeninFigure8andFigure9below.Thisoffersthedevelopermoregranularcontrolovertheflowofdataintheapplication.There

aremanyreasonswhyaDevelopermightchoosetousefunctionsoverExtensions,butafewcommonusecasesare:

?APIcallsneedtobemadeatanotherlayeroftheapplicationstack,outsideofthedirectagentarchitectureflow(e.g.amiddlewaresystem,afrontendframework,etc.)

?SecurityorAuthenticationrestrictionsthatpreventtheagentfromcallinganAPIdirectly(e.gAPIisnotexposedtotheinternet,ornon-accessiblebyagentinfrastructure)

?Timingororder-of-operationsconstraintsthatpreventtheagentfrommakingAPIcallsinreal-time.(i.e.batchoperations,human-in-the-loopreview,etc.)

September202419

Agents

?AdditionaldatatransformationlogicneedstobeappliedtotheAPIResponsethatthe

agentcannotperform.Forexample,consideranAPIendpointthatdoesn’tprovidea

filteringmechanismforlimitingthenumberofresultsreturned.UsingFunctionsonthe

client-sideprovidesthedeveloperadditionalopportunitiestomakethesetransformations.

?Thedeveloperwantstoiterateonagentdevelopmentwithoutdeployingadditional

infrastructurefortheAPIendpoints(i.e.FunctionCallingcanactlike“stubbing”ofAPIs)

WhilethedifferenceininternalarchitecturebetweenthetwoapproachesissubtleasseeninFigure8,theadditionalcontrolanddecoupleddependencyonexternalinfrastructuremakesFunctionCallinganappealingoptionfortheDeveloper.

Figure8