Introductionto

Agents

G.e

Authors:AlanBlount,AntonioGulli,ShubhamSaboo,

MichaelZimmermann,andVladimirVuskovic

IntroductiontoAgentsandAgentarchitectures

Acknowledgements

Contentcontributors

EnriqueChan

MikeClark

DerekEgan

AnantNawalgariaKanchanaPatlollaJuliaWiesinger

Curatorsandeditors

AnantNawalgariaKanchanaPatlolla

Designer

MichaelLanning

November20252

Tableofcontents

FromPredictiveAItoAutonomousAgents 6

IntroductiontoAIAgents· 8

TheAgenticProblem-SolvingProcess ·10

ATaxonomyofAgenticSystems ·14

Level0:TheCoreReasoningSystem 15

Level1:TheConnectedProblem-Solver ·15

Level2:TheStrategicProblem-Solver· ·16

Level3:TheCollaborativeMulti-AgentSystem 17

Level4:TheSelf-EvolvingSystem ·18

CoreAgentArchitecture:Model,Tools,andOrchestration ·19

Model:The“Brain”ofyourAIAgent ·19

Tools:The"Hands"ofyourAIAgent 20

RetrievingInformation:GroundinginReality 21

ExecutingActions:ChangingtheWorld· ·21

FunctionCalling:ConnectingToolstoyourAgent ·22

Tableofcontents

TheOrchestrationLayer 22

CoreDesignChoices 23

InstructwithDomainKnowledgeandPersona· 23

AugmentwithContext. ·24

Multi-AgentSystemsandDesignPatterns 24

AgentDeploymentandServices ·26

AgentOps:AStructuredApproachtotheUnpredictable ·27

MeasureWhatMatters:InstrumentingSuccessLikeanA/BExperiment 29

QualityInsteadofPass/Fail:UsingaLMJudge ·29

Metrics-DrivenDevelopment:YourGo/No-GoforDeployment ·30

DebugwithOpenTelemetryTraces:Answering"Why?" ·30

CherishHumanFeedback:GuidingYourAutomation 31

AgentInteroperability ·31

AgentsandHumans 32

AgentsandAgents 33

AgentsandMoney34

Tableofcontents

SecuringaSingleAgent:TheTrustTrade-Off 34

AgentIdentity:ANewClassofPrincipal ·35

PoliciestoConstrainAccess 37

SecuringanADKAgent 37

ScalingUpfromaSingleAgenttoanEnterpriseFleet 39

SecurityandPrivacy:HardeningtheAgenticFrontier 40

AgentGovernance:AControlPlaneinsteadofSprawl 40

Howagentsevolveandlearn 42

Howagentslearnandselfevolve 43

SimulationandAgentGym-thenextfrontier ·46

Examplesofadvancedagents 47

GoogleCo-Scientist 47

AlphaEvolveAgent 49

Conclusion ·51

Endnotes 52

IntroductiontoAgentsandAgentarchitectures

AgentsarethenaturalevolutionofLanguageModels,madeusefulinsoftware.

FromPredictiveAItoAutonomousAgents

Artificialintelligenceischanging.Foryears,thefocushasbeenonmodelsthatexcelat

passive,discretetasks:answeringaquestion,translatingtext,orgeneratinganimagefromaprompt.Thisparadigm,whilepowerful,requiresconstanthumandirectionforeverystep.We'renowseeingaparadigmshift,movingfromAIthatjustpredictsorcreatescontenttoanewclassofsoftwarecapableofautonomousproblem-solvingandtaskexecution.

ThisnewfrontierisbuiltaroundAIagents.AnagentisnotsimplyanAImodelinastatic

workflow;it'sacompleteapplication,makingplansandtakingactionstoachievegoals.It

combinesaLanguageModel's(LM)abilitytoreasonwiththepracticalabilitytoact,allowing

November20256

IntroductiontoAgentsandAgentarchitectures

November20257

ittohandlecomplex,multi-steptasksthatamodelalonecannot.Thecriticalcapabilityisthatagentscanworkontheirown,figuringoutthenextstepsneededtoreachagoalwithoutapersonguidingthemateveryturn.

Thisdocumentisthefirstinafive-partseries,actingasaformalguideforthedevelopers,architects,andproductleaderstransitioningfromproofs-of-concepttorobust,

production-gradeagenticsystems.Whilebuildingasimpleprototypeisstraightforward,ensuringsecurity,qualityandreliabilityisasignificantchallenge.Thispaperprovidesacomprehensivefoundation:

?CoreAnatomy:Deconstructinganagentintoitsthreeessentialcomponents:thereasoningModel,actionableTools,andthegoverningOrchestrationLayer.

?ATaxonomyofCapabilities:Classifyingagentsfromsimple,connectedproblem-solverstocomplex,collaborativemulti-agentsystems.

?ArchitecturalDesign:Divingintothepracticaldesignconsiderationsforeachcomponent,frommodelselectiontotoolimplementation.

?BuildingforProduction:EstablishingtheAgentOpsdisciplineneededtoevaluate,debug,secure,andscaleagenticsystemsfromasingleinstancetoafleetwith

enterprisegovernance.

Buildingontheprevious

Agentswhitepaper

1and

AgentCompanion

2;thisguideprovides

thefoundationalconceptsandstrategicframeworksyouwillneedtosuccessfullybuild,

deploy,andmanagethisnewgenerationofintelligentapplicationswhichcanreason,actandobservetoaccomplish

goals

3.

IntroductiontoAgentsandAgentarchitectures

November20258

WordsareinsufficienttodescribehowhumansinteractwithAl.Wetendto

anthropomorphizeandusehumantermslike“think”and“reason”and“know.”WedonItyethavewordsfor"knowwithsemanticmeaning"vs"knowwithhighprobabilityof

maximizingarewardfunction."Thosearetwodifferenttypesofknowing/buttheresultsarethesame99.X%ofthetime.

IntroductiontoAIAgents

Inthesimplestterms,anAIAgentcanbedefinedasthecombinationofmodels,tools,anorchestrationlayer,andruntimeserviceswhichusestheLMinalooptoaccomplishagoal.Thesefourelementsformtheessentialarchitectureofanyautonomoussystem.

?TheModel(The"Brain"):Thecorelanguagemodel(LM)orfoundationmodelthatservesastheagent'scentralreasoningenginetoprocessinformation,evaluateoptions,and

makedecisions.Thetypeofmodel(general-purpose,fine-tuned,ormultimodal)dictatestheagent'scognitivecapabilities.AnagenticsystemistheultimatecuratoroftheinputcontextwindowtheLM.

?Tools(The"Hands"):Thesemechanismsconnecttheagent'sreasoningtotheoutside

world,enablingactionsbeyondtextgeneration.TheyincludeAPIextensions,code

functions,anddatastores(likedatabasesorvectorstores)foraccessingreal-time,factualinformation.AnagenticsystemallowsaLMtoplanwhichtoolstouse,executesthetool,andputsthetoolresultsintotheinputcontextwindowofthenextLMcall.

?TheOrchestrationLayer(The"NervousSystem"):Thegoverningprocessthat

managestheagent'soperationalloop.Ithandlesplanning,memory(state),andreasoningstrategyexecution.Thislayerusespromptingframeworksandreasoningtechniques(like

IntroductiontoAgentsandAgentarchitectures

November20259

Chain-of-Thought

4or

ReAct

5)tobreakdowncomplexgoalsintostepsanddecidewhentothinkversususeatool.Thislayerisalsoresponsibleforgivingagentsthememory

to"remember."

?Deployment(The"BodyandLegs"):Whilebuildinganagentonalaptopiseffectiveforprototyping,productiondeploymentiswhatmakesitareliableandaccessibleservice.

Thisinvolveshostingtheagentonasecure,scalableserverandintegratingitwith

essentialproductionservicesformonitoring,logging,andmanagement.Oncedeployed,theagentcanbeaccessedbyusersthroughagraphicalinterfaceorprogrammaticallybyotheragentsviaanAgent-to-Agent(A2A)API.

Attheendoftheday,buildingagenerativeAIagentisanewwaytodevelopsolutionsto

solvetasks.Thetraditionaldeveloperactsasa"bricklayer,"preciselydefiningeverylogicalstep.Theagentdeveloper,incontrast,ismorelikeadirector.Insteadofwritingexplicitcodeforeveryaction,yousetthescene(theguidinginstructionsandprompts),selectthecast

(thetoolsandAPIs),andprovidethenecessarycontext(thedata).Theprimarytaskbecomesguidingthisautonomous"actor"todelivertheintendedperformance.

You'llquicklyfindthatanLM'sgreateststrength—itsincredibleflexibility—isalsoyourbiggestheadache.Alargelanguagemodel'scapacitytodoanythingmakesitdifficulttocompelittodoonespecificthingreliablyandperfectly.Whatweusedtocall“promptengineering”andnowcall“contextengineering”guidesLMstogeneratethedesiredoutput.Foranysingle

calltoaLM,weinputourinstructions,facts,availabletoolstocall,examples,sessionhistory,userprofile,etc–fillingthecontextwindowwithjusttherightinformationtogettheoutputsweneed.AgentsaresoftwarewhichmanagetheinputsofLMstogetworkdone.

Debuggingbecomesessentialwhenissuesarise."AgentOps"essentiallyredefinesthe

familiarcycleofmeasurement,analysis,andsystemoptimization.Throughtracesandlogs,youcanmonitortheagent's"thoughtprocess"toidentifydeviationsfromtheintended

executionpath.Asmodelsevolveandframeworksimprove,thedeveloper'sroleistofurnish

IntroductiontoAgentsandAgentarchitectures

November202510

criticalcomponents:domainexpertise,adefinedpersonality,andseamlessintegration

withthetoolsnecessaryforpracticaltaskcompletion.It'scrucialtorememberthat

comprehensiveevaluationsandassessmentsoftenoutweightheinitialprompt'sinfluence.

Whenanagentispreciselyconfiguredwithclearinstructions,reliabletools,andan

integratedcontextservingasmemory,agreatuserinterface,theabilitytoplanandproblemsolve,andgeneralworldknowledge,ittranscendsthenotionofmere"workflowautomation."Itbeginstofunctionasacollaborativeentity:ahighlyefficient,uniquelyadaptable,and

remarkablycapablenewmemberofyourteam.

Inessence,anagentisasystemdedicatedtotheartofcontextwindowcuration.It

isarelentlessloopofassemblingcontext,promptingthemodel,observingtheresult,

andthenre-assemblingacontextforthenextstep.Thecontextmayincludesystem

instructions,userinput,sessionhistory,longtermmemories,groundingknowledgefromauthoritativesources,whattoolscouldbeused,andtheresultsoftoolsalreadyinvoked.Thissophisticatedmanagementofthemodel'sattentionallowsitsreasoningcapabilitiestoproblemsolvefornovelcircumstancesandaccomplishobjectives.

TheAgenticProblem-SolvingProcess

WehavedefinedanAIagentasacomplete,goal-orientedapplicationthatintegratesareasoningmodel,actionabletools,andagoverningorchestrationlayer.Ashortversionis“LMsinaloopwithtoolstoaccomplishanobjective.”

Buthowdoesthissystemactuallywork?Whatdoesanagentdofromthemomentitreceivesarequesttothemomentitdeliversaresult?

IntroductiontoAgentsandAgentarchitectures

November202511

Atitscore,anagentoperatesonacontinuous,cyclicalprocesstoachieveitsobjectives.

Whilethisloopcanbecomehighlycomplex,itcanbebrokendownintofivefundamentalstepsasdiscussedindetailinthebookAgenticSystemDesign:6

1.GettheMission:Theprocessisinitiatedbyaspecific,high-levelgoal.Thismissionisprovidedbyauser(e.g.,"Organizemyteam'stravelfortheupcomingconference")oranautomatedtrigger(e.g.,"Anewhigh-prioritycustomertickethasarrived").

2.ScantheScene:Theagentperceivesitsenvironmenttogathercontext.Thisinvolvestheorchestrationlayeraccessingitsavailableresources:"Whatdoestheuser'srequestsay?","Whatinformationisinmytermmemory?DidIalreadytrytodothistask?Didtheusergivemeguidancelastweek?","WhatcanIaccessfrommytools,likecalendars,

databases,orAPIs?"

3.ThinkItThrough:Thisistheagent'score"think"loop,drivenbythereasoningmodel.TheagentanalyzestheMission(Step1)againsttheScene(Step2)anddevisesaplan.This

isn'tasinglethought,butoftenachainofreasoning:"Tobooktravel,Ifirstneedtoknowwhoisontheteam.Iwillusetheget_team_rostertool.ThenIwillneedtochecktheiravailabilityviathecalendar_api."

4.TakeAction:Theorchestrationlayerexecutesthefirstconcretestepoftheplan.Itselectsandinvokestheappropriatetool—callinganAPI,runningacodefunction,orqueryingadatabase.Thisistheagentactingontheworldbeyonditsown

internalreasoning.

5.ObserveandIterate:Theagentobservestheoutcomeofitsaction.Theget_

team_rostertoolreturnsalistoffivenames.Thisnewinformationisaddedtotheagent'scontextor"memory."Theloopthenrepeats,returningtoStep3:"NowthatIhavetheroster,mynextstepistocheckthecalendarforthesefivepeople.Iwillusethecalendar_api."

IntroductiontoAgentsandAgentarchitectures

November202512

This"Think,Act,Observe"cyclecontinues-managedbytheOrchestrationLayer,reasonedbytheModel,andexecutedbytheToolsuntiltheagent'sinternalplaniscompleteandtheinitialMissionisachieved.

Figure1:AgenticAIproblem-solvingprocess

Let’stakeareal-worldexampleofhowaCustomerSupportAgentwouldoperateinthis5-stepcycle:

Imagineauserasks,"Whereismyorder#12345?"

IntroductiontoAgentsandAgentarchitectures

November202513

Insteadofimmediatelyacting,theagentfirstentersits"ThinkItThrough"phasetodeviseacompletestrategy.Itreasons:

"Theuserwantsadeliverystatus.Toprovideacompleteanswer/lneedamulti–stepplan:

1.Identify:lmustfirstfindtheorderinourinternaldatabasetoconfirmitexistsandgetitsdetails.

2.Track:Fromthosedetails/lIllneedtoextracttheshippingcarrierIstrackingnumber.Then/lIllquerytheexternalcarrierIsAPlforthelivestatus.

3.Report:Finally/lmustsynthesizetheinformationlIvegatheredintoaclear/helpfulresponsefortheuser."

Withthismulti-stepplaninmind,theagentbeginsexecution.

Initsfirst"Act"phase,itexecutessteponeofitsplan,callingthefind_order("12345")tool.Itobservestheresult—afullorderrecord,includingthetrackingnumber"ZYX987."

Theagent'sorchestrationlayerrecognizesthatthefirstpartofitsplaniscompleteandimmediatelyproceedstothesecond.Itactsbycallingtheget_shipping_

status("ZYX987")tool.Itobservesthenewresult:"OutforDelivery."

Finally,havingsuccessfullyexecutedthedata-gatheringstagesofitsplan,theagent

movestothe"Report"step.Itperceivesithasallthenecessarycomponents,plansthefinalmessage,andactsbygeneratingtheresponse:"Yourorder#12345is'OutforDelivery'!

IntroductiontoAgentsandAgentarchitectures

November202514

ATaxonomyofAgenticSystems

Understandingthe5-stepoperationalloopisthefirstpartofthepuzzle.Thesecondis

recognizingthatthisloopcanbescaledincomplexitytocreatedifferentclassesofagents.Foranarchitectorproductleader,akeyinitialdecisionisscopingwhatkindofagenttobuild.

Wecanclassifyagenticsystemsintoafewbroadlevels,eachbuildingonthecapabilitiesofthelast.

Figure2:Agenticsystemin5steps

IntroductiontoAgentsandAgentarchitectures

November202515

Level0:TheCoreReasoningSystem

Beforewecanhaveanagent,wemuststartwiththe"Brain"initsmostbasicform:the

reasoningengineitself.Inthisconfiguration,aLanguageModel(LM)operatesinisolation,respondingsolelybasedonitsvastpre-trainedknowledgewithoutanytools,memory,orinteractionwiththeliveenvironment.

Itsstrengthliesinthisextensivetraining,allowingittoexplainestablishedconceptsandplanhowtoapproachsolvingaproblemwithgreatdepth.Thetrade-offisacompletelackofreal-timeawareness;itisfunctionally"blind"toanyeventorfactoutsideitstrainingdata.

Forinstance,itcanexplaintherulesofprofessionalbaseballandthecompletehistoryoftheNewYorkYankees.Butifyouask,"WhatwasthefinalscoreoftheYankeesgamelastnight?",itwouldbeunabletoanswer.Thatgameisaspecific,real-worldeventthathappenedafteritstrainingdatawascollected,sotheinformationsimplydoesn'texistinitsknowledge.

Level1:TheConnectedProblem-Solver

Atthislevel,thereasoningenginebecomesafunctionalagentbyconnectingtoandutilizingexternaltools-the"Hands"componentofourarchitecture.Itsproblem-solvingisnolongerconfinedtoitsstatic,pre-trainedknowledge.

Usingthe5-steploop,theagentcannowanswerourpreviousquestion.Giventhe"Mission":"WhatwasthefinalscoreoftheYankeesgamelastnight?",its"Think"steprecognizesthisasareal-timedataneed.Its"Act"steptheninvokesatool,likeaGoogleSearchAPIwiththeproperdateandsearchterms.It"Observes"thesearchresult(e.g.,"Yankeeswon5-3"),andsynthesizesthatfactintoafinalanswer.

IntroductiontoAgentsandAgentarchitectures

November202516

Thisfundamentalabilitytointeractwiththeworld-whetherusingasearchtoolforascore,afinancialAPIforalivestockprice,oradatabaseviaRetrieval-AugmentedGeneration(RAG)isthecorecapabilityofaLevel1agent.

Level2:TheStrategicProblem-Solver

Level2marksasignificantexpansionincapability,movingfromexecutingsimpletaskstostrategicallyplanningcomplex,multi-partgoals.Thekeyskillthatemergeshereiscontextengineering:theagent'sabilitytoactivelyselect,package,andmanagethemostrelevantinformationforeachstepofitsplan.

Anagent'saccuracydependsonafocused,high-qualitycontext.Contextengineering

curatesthemodel'slimitedattentiontopreventoverloadandensureefficientperformance.

Forinstance,considerthe"Mission":"Findagoodcoffeeshophalfwaybetweenmyofficeat1600AmphitheatreParkway,MountainView,andmyclient'sofficeat1MarketSt,

SanFrancisco."

ALevel2agentwillstartcreatingaplan:

1.Think:"Imustfirstfindthehalfwaypoint."

?Act:CalltheMapstoolwithbothaddresses.

?Observe:"ThehalfwaypointisMillbrae,CA."

2.Think:"NowImustfindcoffeeshopsinMillbrae.Theuseraskedfor'good'ones,soIwillsearchforplaceswitha4-starratingorhigher."

IntroductiontoAgentsandAgentarchitectures

November202517

?Act:Callthegoogle_placestoolwithquery="coffeeshopinMillbrae,CA",min_rating=4.0.(Thisiscontextengineering-itautomaticallycreatedanew,focusedsearchqueryfromthepreviousstep'soutput).

?Observe:"Thesearchreturns'MillbraeCoffee'and'TheDailyGrind'."

3.Think:"Iwillsynthesizetheseresultsandpresentthemtotheuser."

Thisstrategicplanningalsoenablesproactiveassistance,likeanagentthatreadsalong

flightconfirmationemail,engineersthekeycontext(flightnumber,date),andactsbyaddingittoyourcalendar.

Level3:TheCollaborativeMulti-AgentSystem

Atthehighestlevel,theparadigmshiftsentirely.Wemoveawayfrombuildingasingle,all-powerful"super-agent"andtowarda"teamofspecialists"workinginconcert,amodelthatdirectlymirrorsahumanorganization.Thesystem'scollectivestrengthliesinthisdivisionoflabor.

Here,agentstreatotheragentsastools.Imaginea"ProjectManager"agentreceivinga"Mission":"Launchournew'Solaris'headphones."

TheProjectManageragentdoesn'tdotheentireworkitself.ItActsbycreatingnewMissionsforitsteamofspecializedagentsmuchlikehowitworksinthereallife:

1.DelegatestoMarketResearchAgent:"Analyzecompetitorpricingfornoise-cancelingheadphones.Returnasummarydocumentbytomorrow."

2.DelegatestoMarketingAgent:"Draftthreeversionsofapressreleaseusingthe'Solaris'productspecsheetascontext."

IntroductiontoAgentsandAgentarchitectures

November202518

3.DelegatestoWebDevAgent:"GeneratethenewproductpageHTMLbasedontheattacheddesignmockups."

Thiscollaborativemodel,whilecurrentlyconstrainedbythereasoninglimitationsoftoday'sLMs,representsthefrontierofautomatingentire,complexbusinessworkflowsfromstarttofinish.

Level4:TheSelf-EvolvingSystem

Level4representsaprofoundleapfromdelegationtoautonomouscreationandadaptation.Atthislevel,anagenticsystemcanidentifygapsinitsowncapabilitiesanddynamically

createnewtoolsorevennewagentstofillthem.Itmovesfromusingafixedsetofresourcestoactivelyexpandingthem.

Followingourexample,the"ProjectManager"agent,taskedwiththe'Solaris'launch,mightrealizeitneedstomonitorsocialmediasentiment,butnosuchtooloragentexistson

itsteam.

1.Think(Meta-Reasoning):"Imusttracksocialmediabuzzfor'Solaris,'butIlackthecapability."

2.Act(AutonomousCreation):Insteadoffailing,itinvokesahigh-levelAgentCreatortoolwithanewmission:"Buildanewagentthatmonitorssocialmediaforkeywords'Solarisheadphones',performssentimentanalysis,andreportsadailysummary."

3.Observe:Anew,specializedSentimentAnalysisAgentiscreated,tested,andaddedtotheteamonthefly,readytocontributetotheoriginalmission.

Thislevelofautonomy,whereasystemcandynamicallyexpanditsowncapabilities,turnsateamofagentsintoatrulylearningandevolvingorganization.

IntroductiontoAgentsandAgentarchitectures

November202519

CoreAgentArchitecture:Model,Tools,andOrchestration

Weknowwhatanagentdoesandhowitcanscale.Buthowdoweactuallybuildit?

Thetransitionfromconcepttocodeliesinthespecificarchitecturaldesignofitsthreecorecomponents.

Model:The“Brain”ofyourAIAgent

TheLMisthereasoningcoreofyouragent,anditsselectionisacriticalarchitecturaldecisionthatdictatesyouragent'scognitivecapabilities,operationalcost,andspeed.

However,treatingthischoiceasasimplematterofpickingthemodelwiththehighestbenchmarkscoreisacommonpathtofailure.Anagent'ssuccessinaproduction

environmentisrarelydeterminedbygenericacademicbenchmarks.

Real-worldsuccessdemandsamodelthatexcelsatagenticfundamentals:superior

reasoningtonavigatecomplex,multi-stepproblemsandreliabletoolusetointeractwiththe

world

7.

Todothiswell,startbydefiningthebusinessproblem,thentestmodelsagainstmetricsthatdirectlymaptothatoutcome.Ifyouragentneedstowritecode,testitonyourprivatecodebase.Ifitprocessesinsuranceclaims,evaluateitsabilitytoextractinformation

fromyourspecificdocumentformats.Thisanalysismustthenbecross-referencedwiththepracticalitiesofcostandlatency.The"best"modelistheonethatsitsattheoptimalintersectionofquality,speed,andpriceforyour

specifictask

8.

IntroductiontoAgentsandAgentarchitectures

November202520

Youmaychoosemorethanonemodel,a"teamofspecialists."Youdon'tuseasledgehammertocrackanut.ArobustagentarchitecturemightuseafrontiermodellikeGemini2.5Profortheheavyliftingofinitialplanningandcomplexreasoning,butthenintelligentlyroutesimpler,high-volumetasks—likeclassifyinguserintentorsummarizingtext—toamuchfasterand

morecost-effectivemodellikeGemini2.5Flash.Modelroutingmightbeautomaticorhard-codedbutisakeystrategyforoptimizingboth

performanceandcost

9.

Thesameprincipleappliestohandlingdiversedatatypes.Whileanativelymultimodal

modellike

Geminilivemode

10offersastreamlinedpathtoprocessingimagesandaudio,

analternativeistousespecializedtoolslikethe

CloudVisionAPI

11or

Speech-to-TextAPI

12.Inthispattern,theworldisfirstconvertedtotext,whichisthenpassedtoalanguage-onlymodelforreasoning.Thisaddsflexibilityandallowsforbest-of-breedcomponents,butalsointroducessignificantcomplexity.

Finally,theAIlandscapeisinastateofconstant,rapidevolution.Themodelyouchoosetodaywillbesupersededinsixmonths.A"setitandforgetit"mindsetisunsustainable.

Buildingforthisrealitymeansinvestinginanimbleoperationalframework—an"AgentOps"

practice

13.WitharobustCI/CDpipelinethatcontinuouslyevaluatesnewmodelsagainsty