監(jiān)管智能體：Agentic+Supervision的未來研究報告+The+Future+of+Agentic+Supervision

ofAgentic

Supervision

TheFuture

ΛRFΛCT

AIISABOUTPEOPLE

WEACCELERATEDATAANDAIADOPTIONTOPOSITIVELYIMPACT

PEOPLEANDORGANIZATIONS.

25

COUNTRIES

1700

EMPLOYEES

+1000

CLIENTS

Artefactisagloballeaderinconsultingservices,specializedindatatransformation

anddata&digitalmarketing,fromstrategytothedeploymentofAIsolutions.

Weareofferingauniquecombinationofinnovation(Art)anddatascience(Fact).

STRATEGY&TRANSFORMATION|AIACCELERATION|DATAFOUNDATIONS&BI

IT&DATAPLATFORMS|MARKETINGDATA&DIGITAL

Executivesummary

LastFebruary,wepublished“TheFutureofWorkwithAI”,ourfirststudyonAgenticAI.WefoundthatalthoughAIagentswillreplacehumansontediousandrepetitivetasks,anewtypeofworkwillappear:AgenticSupervision.Duringtheindustrialrevolution,machinesreplacedhumansonmanualtasks,butnewjobsappearedsuchasmachinepurchasing,operationalsupervisionandmaintenance.WithAgenticAI,cognitivejobswillbereplacedbyotherhigher-levelandmoreproductivecognitivejobs.ThisstudyintendstodeepdiveintotheearlydaysofAgenticSupervisionandtodrawtheoutlineoftheFutureofSupervisionintermsofAgentlifecyclemanagement,governanceandsupervisiontooling.

TogatherthecurrentstateofAgenticSupervision,wein-terviewed14enterprisesand5ArtefactAgenticProductManagers&Engineers.WealsocontactedkeyAgenticSupervisionproviders,includingmajorData&AIplatformswithyearsofsoftwaresupervisionexperience(suchasGoogleandMicrosoft)aswellasspecializedstart-ups(WB,Giskard,RobustIntelligence…).

ThefirstinsightwefoundisthatwhileAgenticSupervisionextendstheprinciplesestablishedinDevOps(softwareop-erations),DataOps(dataoperations),andMLOps(MachineLearningoperations),itdramaticallyincreasesthedemandforrobustgovernancetokeepAIAgentsalignedandundercontrol.Indeed,with“softwarethatstartstothink”,unseenrisksareemerging,suchashallucination,reasoningerrors,inappropriatetone,intellectualpropertyinfringementorevenpromptjacking.Mitigatingthesereliability,behavioral,regulatoryandsecurityrisksnowrequiresgovernancethatisnotonlymorerigorousbutalsobroaderthanwhathaspreviouslybeenappliedtotechproducts.

Thismarkedlygreaterneedforgovernanceisthechal-lengethatmaydefinetheemergingoperationalparadigmof“AgentOps”.Interestingly,AgentOpswillneedtobuilduponeachorganization’sexistingDevOps,DataOps,andMLOpsfoundationsandgovernance,andcompanieslag-

RFΛCT

THEFUTUREOFAGENTICSUPERVISION

“WefoundthatalthoughAIagentswillreplacehumansontediousandrepetitivetasks,anewtypeofworkwillappear:

AgenticSupervision.”

gingintheseoperationaldomainswillhavetobridgeanygapsintheseareaswhilesettingtheirAgenticgovernanceframework.

Thesecondmajorchallengeidentifiedbyourinterview-eesistheneedtostrengthentheirAIsupervisiontooling.ManyarecurrentlyrelyingonexistingRPAandDev/Data/MLOpstools,orexperimentingwithcustom-builtsolutionsastheysearchformoresustainable,long-termoptions.Theabundanceofearly-stagetoolsandtheneedtoenvisionacohesive,end-to-endsupervisionsystemthatintegratesmultiplecomponents,promptedustoexplorethetechno-logicaldimensionsofagenticsupervisioningreaterdepth.AswithanyTechOpsframework,AgentOpssupervisioninvolvesthreefundamentalstages:(1)Observe,(2)Evaluate,and(3)Monitorandmanageincidents.Whilethethirdstagerepresentsthelargestsupervisioneffortandtime,thefirsttwoareessentialtoensuringeffectiveriskmanagement.Withnewcategoriesofriskstomonitorandconsequently,newlogs,traces,andevaluationmechanismstoestablish,it’sclearwhyintervieweesconsistentlyemphasizedtheneedfortherighttoolstosupportscalableandreliablesupervision.

3

EXECUTIVESUMMARYTHEFUTUREOFAGENTICSUPERVISION

“Supervisionshouldnotbeanafterthought,itmustbe

embeddedearlyintheagent’sdesignanddevelopment.”

Ourresearchintoagenticsupervisiontoolsrevealedthreekeyinsights.First,thereiscurrentlynoall-in-onesolutionavailable.MajorcloudproviderslikeGoogleandMicrosoftareactivelydevelopingandreleasingsupervisiontoolsandframeworksaimedatcoveringthefullspectrumofsupervisionneedsforteamsbuildingagentsonplatformssuchasVertexAI(Google)andCopilotStudio(Microsoft).Second,agentsupervisionfallsintotwocategories:pro-activeandreactive.Proactivesupervisionisappliedduringdevelopmenttotestagentsagainstdefinedscenariosor,inproduction,tocontinuouslyguardagainstemergingthreats,particularlyintheareaofsecurity,ortocollectaggregatedperformancedata.Itsgoalistoimproveagentbehaviorovertime.Reactivesupervision,ontheotherhand,focusesondetectingandhandlingliveincidents.Althoughbothtypesrelyonobservabilitytoolsandmayusesimilarevaluationmechanisms,theydiffersignificantlyindatasources,eval-uationgranularity,andresponsestrategies.Finally,ourthirdinsightisthatagenticobservability,evaluation,andriskmitigationremaincomplexandrapidlyevolvingdomains.Weanticipatesubstantialadvancementsinsupervisiontoolingoverthecomingyears.

Eachphaseoftheagenticsupervisioncycle;observe,evaluate,andsupervise,presentsitsownsetofchal-lenges.

Observabilityfirstrequiresanticipatingwhatdatatocapture,whichdependsheavilyonhavingaclearlydefinedevaluationandsupervisionstrategy.Withoutthisforesight,teamsriskeithercollectingtoolittleinformationorbeingoverwhelmedbyvast,unstructuredtracesthathindermanualrootcause

analysis.ToolslikeLangSmithandLangChainareincreas-inglyusedtostructureandstreamlinetheobservationofagentbehavior.AnothermajorchallengeliesintheopacityofLLMreasoning,whichmustbecounteredbydeliberatelydesigningagentarchitecturesandworkflowstoensuretraceabilityandtransparency.

EvaluationinagenticAIissignificantlymorecomplexthanintraditionalsoftwareordataqualityassessments.Wheredeterministictestsbasedonobservabilityqueriesaresuf-ficientinclassicalDevOpsandDataOps,agenticsystemsoftenrequireAItoevaluateAI.ThishasledtotheriseofLLM-as-a-judgetechniques;acounterintuitiveapproachwhereonemodelassessestheoutputofanother.Whilethisraisesconcerns(whytrustflawedAItojudgeflawedAI?),studiesshowitoftenproducesmoreconsistentandscalableresultsthanhumanreviewers.Nonetheless,acommonpainpointamongintervieweeswasthedifficultyofbuildingreliablegroundtruthdatasets,expert-curatedquestion-answerpairs,tobenchmarkagentresponses.Humanevaluatorstendtodisagreeandoftenlackcom-pletenessintheiranswers.

Finally,supervisionandmitigationfacechallengesaroundprioritization.Withagrowingnumberofmetricsandalerts,teamscanquicklybecomeoverwhelmed.Standardizedframeworksforalertingandmetricmanagementareamusttobringstructureandclaritytoagenticsupervision.

Onlyahandfuloforganizationshavesuccessfullyestab-lishedeffectivegovernanceandstandardsforagenticAI.Thosewithmaturesoftwareanddatagovernanceframe-

4ΛRFCT

EXECUTIVESUMMARY

“AgenticSupervisionis

theFutureofWorkwithAI!”

workshavehadaheadstart,benefitingfromstrongfoun-dationsandawell-establishedcultureofobservabilityandsupervision.Weobservedthatleveragingexistingsoftware,RPA,anddatasupervisionpractices,processes,andtoolscansignificantlyaccelerateprogress.However,thekeychal-lengeliesinadaptingthesetothedynamicrisksandevolvingtoolsetsspecifictoagenticAI,andinbuildingadedicated,future-readygovernanceframework.Relyingtoolongonlegacyapproaches,includingdeterministiclogicandcus-tom-builttools,canbecomeaconstraint,limitingteamstonarrow,tightlycontrolledagenticworkflowsandpreventingtheadoptionofmoreautonomous,AI-orchestratedagents.

Allintervieweesemphasizedthatthekeytoeffectiveagenticsupervisionisanticipation.Supervisionshouldnotbeanaf-terthought,itmustbeembeddedearlyintheagent’sdesignanddevelopment.Settingupobservabilityandevaluationmechanismsonlyoncetheagentisinproductionistoolate.Identifyingflawsatthatstageoftenmeansreworkingtheentireagent,whichisfarmorecostlythaninvestinginrobustsupervisionfromthestart.

Thegoodnewsisthatavarietyoftestedtoolcombinationsandemergingagenticframeworksarealreadyavailable.WestronglyrecommendthatenterpriseAIgovernanceteamsdefinetheirownstandardizedframeworkandtoolsettobeappliedacrossallagenticdevelopment.Thisbecomesevenmorecriticalasagentsbegintointerconnect,makingsys-tem-widecontrolandsupervisioninteroperabilityessential.

Tosucceed,AIgovernancemustalsoaligncloselywithstrongITandDataGovernancepractices,sinceagents

RFΛCT

THEFUTUREOFAGENTICSUPERVISION

relyonenterprisedataandITsystemsto‘think’andtake‘a(chǎn)ction.’JustasITanddatagovernancerequiredbusinessinvolvementinthepast,oneofthekeytakeawaysfromourstudyisthatagenticgovernancewilldemandevendeeperbusinessengagement.

Unliketraditionalsoftwareordatasupervision,typicallyhandledbyITordatateams(andinthemostmatureor-ganizations,byabusiness-leddatagovernancenetwork),agentsupervisionwillneedtobebusiness-owned.GiventheinherentunpredictabilityofAIagents,incidentresponsesof-tenrequiredomainexpertise.Asaresult,thebusinessmustbeactivelyinvolvednotjustinmonitoring,butinframingagentbehaviorfromtheoutset.Thisrepresentsasignificantculturalshift:agenticAIblursthelinesbetweenIT,data,andbusiness,andwillrequirenewwaysofworkingbasedoncross-functionalcollaboration.AgenticSupervisionistheFutureofWorkwithAI!

FlorenceBénézit

ExpertPartnerData&AIGovernance

HananOuazan

ManagingPartner,LeadGenerativeAI

5

THANKS&ACKNOWLEDGMENTSTHEFUTUREOFAGENTICSUPERVISION

Methodology

ThisstudyisbasedonaqualitativeresearchapproachdesignedtoexploretheemergingchallengesandgovernancepracticessurroundingtheearlyimplementationsofautonomousAIagentsinorganizations.Bycombiningexpertinterviewswithanin-depthanalysisoftheevolvingtechnologicallandscape,weaimedtomapcurrentpractices,identifyoperationalneeds,andunderstandthevaluepropositionsofavailablesolutionsforagentobservability,evaluation,andsupervision.

Weconducted20+interviewswithprofessionalsdirectlyinvolvedinthedeployment,governance,ortechnicaldevelopmentofagenticsystems.Theseincluded:

—AIandDataLeaders,suchasChiefDataOfficers,HeadsofAI,andDataPlatformDirectors,whosharedtheirstrategicvisiononagentimplementation,riskmanagement,andtheevolutionofdatainfrastructure.

—ProductManagersandInnovationExecutiveswhoofferedinsightsintooperationalusecases,organizationalreadiness,andtheshifttowardagent-centricarchitectures.

—Compliance,Security,andITGovernanceExperts,

whoprovidedcriticalinputonregulatoryexpectations,ethicalrisks,andtheemergingneedforreal-timecontrolmechanismstailoredtoAIagents.

—FoundersandChiefsofScienceofAItoolingcompanies,

whosefeedbackhelpedassessthestateofthemarketacrossthreekeyfunctions:observability,evaluation,andactivesupervisionofAIagents.

Intervieweesrepresentedadiverserangeoforganizations,includingmajorcorporations(insectorssuchasenergy,telecom,pharmaceuticals,andluxury),globaltechplayers,andhigh-growthstartups,ensuringarichandnuancedunderstandingofthetopic.

Inparallel,weconductedasystematicreviewofoveradozentoolsandplatformsofferingcapabilitiesrelevanttoagentgovernanceincludingLangfuse,LangSmith,DeepEval,CopilotStudio,VertexAI,Ragas,Weights&Biases,PRISMEval,DeepEval,RobustIntelligence,Giskard…Eachsolutionwasanalyzedusingadedicatedframeworkthatcross-referencedthreedimensionsofquality(Reliability,BehavioralAlignment,Security)withthreestagesofsupervision(Observation,Evaluation,ActiveSupervision).

Byintegratingreal-worldpractitionerfeedbackwithastructuredtechnologicalbenchmark,thisstudyaimstoofferapragmaticandforward-lookingperspectiveonhowcompaniescanresponsiblyscaleagenticAIsystems.

SpecialThanks&Acknowledgments

ENTERPRISEINTERVIEWEES

YoannBersihand,VPAITechnology,SCHNEIDER

ArthurGarnier,ITChiefofStaff&DataScientist,ARDIANJean-Fran?oisGuilmard,CDO,ACCOR

PaulSaffers,DeputyCDO,VEOLIA

AlexisVaillant,HeadAutomatisation,ORANGE

LeoWang,DataProtectionOfficer,LOUISVUITTONCHINA

AGENTOPSSTACKINTERVIEWEES

AlexCombessie,Co-founder&Co-CEO,GISKARD

SaloméFroment,AccountDirectorFrance,WEIGHTS&BIASESéricHoresnyi,HeadofAIGo-To-Market,GOOGLEFRANCE

AminKarbasi,SeniorDirector,CISCOFOUNDATIONAIRESEARCH(FormerChiefScientistatRobustIntelligence)

Jean-LucLaurent,GenerativeAI/MLSpecialist,GOOGLE

PierrePeigné,Co-founderandChiefScienceOfficer,PRISMEvalChrisVanPelt,Co-founder&CISO,WEIGHTS&BIASES

MarcGardette,DeputyCTO,MICROSOFTFRANCE

6ΛRFCT

TABLEOFCONTENTSTHEFUTUREOFAGENTICSUPERVISION

8

Introduction

9I—AgenticAIrisksareshakingupthetech

governance&supervisiongame.

10AgenticAIorwhensoftwarestartstothink.

14Newtech,oldproblems:whygovernanceisacontinuum.

18Nomorewatchingfromthesidelines:AgenticAIputssupervisioninbusinesshands.

24II—ThenewAgentOpsstack:tests,guardrailsandfeedbackloops.

25Pre-productiontestingmustembracevariabilitytoensureagentreadiness.

35Guardrailsprotectoperationsbymanagingrisksduringagentexecution.

41Agentsupervisionspansfromimmediateruntimeactionstofutureplanningdecisions.

45III—SecureandaccelerateAgenticAIwith

standards&globalgovernance.

46Technicalteamsneedclearstandardstobuildanddeployagentsefficientlyandresponsibly.

50Scalingmulti-agentsystemsrequiressharedprotocolsforinteroperabilityandmanageability.

55BusinessteamsneedtoorganizeglobalAIgovernanceandsupervisionprotocols.

58

Conclusion

RFΛCT7

INTRODUCTIONTHEFUTUREOFAGENTICSUPERVISION

Introduction

If,asshowninourpreviousstudy,thefutureofworkwithAIliesinsupervisingAIagents,thenitisessentialtoensurethatthisnewformofworkbecomesabetterexperiencethanthecognitivetasksitreplaces.Manu-allyoverseeingeverystepanddecisionmadebyanagentwouldquicklybecomeatedious,evenmoredrainingtaskthansolvingtheproblemdirectlyourselves.So,howcanwedobetter?Thisstudyexploreswhat’strulyatstakeinagenticsupervisionandhowearlytoolsarebeginningtoshapewhatthisnewtypeofworkmightlooklike.

Wetakeabroadviewofwhatsupervisionmeans.Itstartswithsettingupautomatedloggingandtracingsystems.Italsoinvolvesdesigningevaluationandalert-ingframeworksthatguidethefinalandmostvisiblestep:takingaction(manuallycorrectingmistakes,relaunchinganagentictaskwithbettercontext,mitigatingincidents,identifyingareasforimprovement,andprioritizingde-velopmentefforts).Supervisingagentsmirrorsmanyaspectsofhumancollaboration:definingjobdescriptions(agentobjectives),recruiting(designinganddeployingnewagents),trainingandcoaching(monitoringandup-

datingbehavior),andongoingcollaboration(providingin-putsandsupporttoagents,butalsolearningfromagentsandthebusinesscontexttheycollectintheirmemory).

Webelievethatthesupervisionofasingleagentwillnotfalltojustoneperson.Agenticsupervisionisinherentlymultidimensional.Forinstance,businessoperationsmayoverseerelevanceandaccuracy;ethicsteams,compli-anceandtone;businessleaders,valueandeconomicviability;andcybersecurityteams,safetyandmaliciousattackriskmitigation.

Thisstudyfocusesonbestpracticesforagenticgov-ernance,supervisionprocesses,andthesupportingtools.Whilethisdomainisstillemergingandlikelytoevolvesignificantly,wealsoobservestrongcontinuitywithestablishedpracticesfromsoftware,RPA,data,andMLsupervision.DespitetheuniquechallengesposedbytheprobabilisticbehaviorofAIagents,manystablefoundationsalreadyexist.Embracingthesefoundationsnowiscriticaltoensuringthesuccessofearlyagenticinitiatives.

GeneratedwithChatGPT

8RFCT

THEFUTUREOFAGENTICSUPERVISION

I

AgenticAIrisksareshakingupthetechgovernance&

supervisiongame.

10

I.A

—AgenticAIorWhenSoftwareStartstoThink.

14I.B—NewTech,OldProblems:WhyGovernanceIsaContinuum.

18I.C—Nomorewatchingfromthesidelines:AgenticAIputssupervisioninbusinesshands.

9

IAGENTICAIRISKSARESHAKINGUPTHETECHGOVERNANCE&THEFUTUREOFAGENTICSUPERVISIONSUPERVISIONGAME.

I.AAgenticAIorWhenSoftwareStartstoThink.

AIagentsradicallydifferfromsoftware:theyareautonomousandgoal-driven.

Traditionalsoftwarefollowspredeterminedlogic,andchat-botsoperatewithinrigidtemplatesanddeterministicdeci-siontrees.Incontrast,agenticAIsystemsgomuchfurther:theyinterpretcontext,planactions,andexecutetasksbychainingdecisionsacrossvarioustoolsandAPIs.Theseagentsdon’tsimplywaitforusercommands,theypursueobjectives,evaluateintermediateoutcomes,andadjusttheirstrategiesonthefly.Thisautonomousreasoningmakesthemfeellessliketoolsandmorelikecollaborators.UnlikeRPAbots(RoboticProcessAutomation)orevenstandalonelargelanguagemodels(LLMs),agenticAIsys-temsaregoal-orientedandtask-complete,builttoachieveanoutcome,notjustfollowinstructionsorgeneratethemostlikelynextresponsetoaprompt.

Thismarksafundamentalshiftinthesoftwaredevelop-mentparadigm.Insteadofhardcodinglogicupfront,youdefinegoalsandsetconstraintsandtheagentautono-mouslyconstructsitsownplan.Itmaychainprompts,callAPIs,search&querydatastores,orevencreatesubgoalsasneeded.Ratherthanfollowingafixedpath,thesystemcontinuouslyadaptsitsactionstowhat’smostlikelyto

succeed.Whilethisopensthedoortomajorproductiv-itygains,italsodisruptstraditionalgovernancemodels:Howdoyoutestasystemwhoseoutputschangewitheveryrun?Howcanyoucontrolbehaviorthatvariesovertime,withoutresortingtoconstanthumanoversightandintervention?

“What’sdifferentwithagentsisthattheydon’tjustfollowascript.Theyinterpretinstructions,decidehowtoachievegoals,andofteninfermorethanyoutoldthemto.Thatopensupanewlayerofunpredictability.You’renotsuper-visingcode,you’resupervisingintent.”

ArthurGRENIER

ITChiefofStaff&SeniorDataScientist

ARDIAN

IAGENTICAIRISKSARESHAKINGUPTHETECHGOVERNANCE&THEFUTUREOFAGENTICSUPERVISIONSUPERVISIONGAME.

AgenticAIcan’tbemade100%predictableandcallsforgovernancereinventiontobalancevalueandrisks.

Thefirstgenerationofautomationtools,includingRPA,macrosandrule-basedbots,offeredpredictabilitybyde-sign.Theymimickeduseractionsstepbystep,withinwell-definedworkflows.EventraditionalMachineLearningsystems,despitetheirinternalcomplexityandprobabilisticnature,operatedwithinclearboundaries:structuredinputsandoutputs.Incontrast,LLMsacceptunstructuredtextinputsandcangenerateawiderangeofoutputs,ofteninunpredictableformats.AgenticAIexacerbatesbehaviorcomplexityevenfurther,agentsnavigatedynamicenviron-ments,drawonmultipleknowledgesources,andadapttheiractionsautonomouslyinrealtime.Theirbehaviorisinfluencednotjustbytrainingdataorpredefinedrules,butbyhumanprompts,toolusage,memorystate,andimplicitknowledgebakedintotheirfoundationmodels.

Legacygovernancemodelsreliedondeterministicin-put-outputcontrol:supplytestdata,verifyresults,tracebugs.Butagenticsystemsblurthatline.Asinglepromptmightleadtohallucinations,multipleAPIcalls,toolinterac-tions,ormemoryrecalls,allpotentiallydifferenteachtime.Thisabstractionbetweenintentandexecutioncreatesagovernancecontrolgapintermsoftechnicalvisibility,pro-cessreadinessandaccountability:rulescanbebypassed,edgecasesoverlooked,andbehavioralregressionsmaygounnoticeduntiltheycauserealissues.

Asaresult,supervisingagentsshiftstheeffortweightfromverifyingcodetoobservingpairsofinputsandoutputs,andpiecingtogethertheirdecision-makingret-rospectively.Asforsoftwareanddatamanagement,thisobservation&analysisefforthappensbothoffline,beforedeploymentongroundtruthorsyntheticdata,andonlineonproductiondata.Allintervieweesstressedtheimportanceofsettingupagenticsupervisionupfronttorigorouslytestagentswhilebeingdevelopedbutalsotoanticipateonlinesupervisionaccountabilityandreme-diationprocesses.

“Unliketraditionalsoftware,AIdevelopmentisfundamentallyprobabilistic.CodeisnolongerthecoreIP,learningis.Whatmattersisknow-ingwhatworks,whatdoesn’t,andwhy.”

ChrisVanPelt

Co-founder&CISO

10ΛRFCTRFΛCT11

IAGENTICAIRISKSARESHAKINGUPTHETECHGOVERNANCE&THEFUTUREOFAGENTICSUPERVISIONSUPERVISIONGAME.

Thisunpredictabilityshiftintroducestheneedforlarge-scale,statisticalvalue&riskevaluation.

Asaconsequenceofthisunpredictability,theemergenceofagenticAIhasintroducedaprofoundcontrolchallenge:traditionalQA(QualityAssessment)methodsarenolongeradequate.Previously,ahandfulofunittestsmatchingfixedinputstotheirexpecteddeterministicoutputswasenoughtovalidatehardcodedlogic.Incontrast,AIagentsnowrequiretestingacrossabroadspectrumofpossibleinputs,witheachtestscenariorigorouslyandrepeatedlyruntoaccountfortheirnon-deterministicbehavior.Ontopofthat,evaluatingtheirperformancemeansinterpretingun-structuredandvariabletextoutputs,whichmakesitmuchhardertoconsistentlydefineandmeasurewhat“quality”reallymeans.Outputqualitymayneedtobeassessedalongmultipledimensions,includingfactualaccuracy,completeness,security,andalignmentwithuserintent.

Oncequalityisassessed,asecondchallengeemerges:identifyingtherootcausesofagentfailurestosupportim-provementormanagerunincidents.Thisrequiresdetailed,transparentloggingoftheagent’sreasoningprocess,accessibletoadiversesetofsupervisingstakeholders;developers,complianceofficers,businessowners,anddomainexpertsalike.

“Theneedtoclosethissupervisionandgovernancegaprisesveryearlyintheenterpriseagenticjourney.”

Theneedtoclosethissupervisionandgovernancegaprisesveryearlyintheenterpriseagenticjourney.Asagenticsystemsbegininterpretingcomplexbusinesscontextsandtakingautonomousdecisions,therisksandresponsibilitiesgrow.Whileagentsarealreadybeingdeployedinenterprisepilotsacrossvariousfunctions,thetechnical,organization-al,andlegalinfrastructuresrequiredforrobustsupervisionremainunderdeveloped.Legacygovernanceframeworksareinsufficientandenterprisesneedtoupgradeitwithanew,test-intense,purpose-builtapproach.

“AftertheDigitalandMobilerevolutions,wearenowenteringathirdwaveofmediadisrup-tion:AIagents.Theseagentswillincreasinglymediateourinteractionswithcompanies,

transforminghowwesearch,learn,shop,

work,andcommunicate.Imaginethatin2030,40%ofinteractionsbetweenconsumersandcompanieswillbeshapedbyAI.Buthowdowecontrolthereliabilityandsecurityrisksoftheseagents?”

AlexCOMBESSIE

Co-founder&Co-CEO

}PGiskard

12ΛRFCT

IAGENTICAIRISKSARESHAKINGUPTHETECHGOVERNANCE&THEFUTUREOFAGENTICSUPERVISIONSUPERVISIONGAME.

TECHNOLOGY

Giskardisanopen-sourcetestingplatformdesignedtoensurethequality,security,andcomplianceofAImodels.Itautomatesthedetectionofvulnerabilitiessuchashallucinations,biases,andsecurityflawsinLLMsandagents.Giskard’sfeaturesincludeautomatedtestgeneration,continuousmonitoring,andcollaborativetoolsthatfacilitatecross-functionalteamworkamongdatascientists,developers,andbusinessstakeholders.

FEATURECOVERAGE

Eliability,Regulatorycompliance,Security,FinOps,Latency

OBSERVE.

Giskarddoesnotofferreal-timeob-servabilityfeaturessuchastrackinglatency,tokenusage,orcostmet-rics.Itsprimaryfocusisonpre-de-ploymenttestingandvulnera