Anthropicengineer's14-steproadmap,usingathree-layerarchitectureandfour-layercompoundingstack,deconstructshowtobuildaself-improvingagentsystemaroundFable5thatcompoundsiteratively;contentsourcedfromengineeringarticlesandtheteam'spublicexperiments.Anthropic'sContinualLearningBenchexperimentshowsFable5withmemoryachieves73%verificationcoverage.##Four-LayerCompoundingStackArchitecture:Layer-by-LayerOperationfromPrimitiveLayertoSelf-ImprovementLayer

(Source:AnthropicFable5)Accordingtotheoriginalframework,thefour-layercompoundingstackisbuiltfromthebottomup.Theoutputofeachlayerflowsupwardtothetoplayer,whereitisscored,distilled,andwrittenbacktothememorylayer:·Layer1(Primitives)includesFable5itself,sub-agents,worktrees,andtools.Thisisthelayermostuserscurrentlyuse;·Layer2(Orchestration)uses/goalandOutcomesforself-correctionloops,multi-steporchestrationwithdynamicworkflows,andRoutinesforlong-runningcloudexecution;·Layer3(Memory)includesstatefiles(STATE.md),Skills,KnowledgeBases,anddistilledlessons;·Layer4(Self-Improvement)includesvisualself-verification,evalloops,andruledistillation.Theagentscoresitsownoutput,refinesSkills,andwriteslessonsbacktomemory,closingtheloop.##/goalvs.Outcomes:UseCaseComparisonforTwoGoal-DrivenLoopsAccordingtoAnthropicengineeringdocuments,/goal(ClaudeCode)andOutcomes(ClaudeManagedAgents)sharethesamecoreshape:anindependentscorerchecksthework,a"notmet"judgmenttriggersthenextiteration,andtheloopexitswhenthescorerpasses.Theselectionrulesforthetwoareasfollows:/goalissuitableforlocal,session-basedtaskswithmeasurableendstates(e.g.,codedebugging,single-filerefinement),usingplaintextgoalsandamodelscorer;OutcomesissuitablefortasksthatneedtorunacrosshoursordaysonAnthropic'shostedinfrastructure(e.g.,MLtraining,long-runningmigrations),usingfile-basedscoringcriteria,asub-agentscorer,andahardmax_iterationslimit.Keystructuralprinciplesharedbyboth:theagentwritingcodeisnottheagentscoringit.##ContinualLearningBenchExperiment:Fable5MemoryVerificationCoverage73%AccordingtoAnthropic'sContinualLearningBench1.0experiment,thefive-stagememoryprogression(Fail→Investigate→Verify→Distill→Consult)showsthefollowingperformancedifferencesacrossmodels:Sonnet4.6:ExitsatStage1,memoryisonlyfailurenotesandunresolvedguesses,rarelyreviewspreviousnotes,memorydoesnotcompoundOpus4.7:ExitsatStage3,buildsreferencedocumentswithuncertaintyannotations,verificationcoverage7-33%(median~17%)Fable5:Tendstocompletetheentirefive-stageprogression,achieving73%verificationcoverageinitsstrongestrun(22outof30questions),anddistillslearningsintogeneralrulesapplicabletofuturetasksAdditionally,intheParameterGolfexperiment,Fable5pairedwithanindependentverifierexploredlargerarchitectural-levelchanges,navigatednegativeintermediateresults,andultimatelyachievedapproximatelysixtimesmoreimprovementsthanOpus4.7.##Five-StageMemoryProgressionandStateFileArchitecture:FiveStructuralSectionsofSTATE.mdAccordingtoAnthropicengineeringdocuments,thefivesectionsofthestatefile(STATE.md)correspondtothefivememorystages:Verifiedfacts(factswhereguessinghasstopped,outputofStage3),Generalrules(distilledrulesbeyondspecificcases,outputofStage4),Openfailures(ongoingStage1-2work),Lessonslearned(moreStage4output),Lastsession(continuationindicatorforStage5).DatafromContinualLearningBenchshowsthatifSTATE.mdandrelatedSkillsarenotreadatthestartofeachsession,evenFable5exhibitsSonnet-levelmemorybehavior.Skillsarestoredin~/.claude/skills/,arecross-projectusable,andserveasthelong-termaccumulationcarrierofproceduralmemory;everyconfirmedlessonshouldbewrittenintoaSkill,notjustSTATE.md.##Fable5SafetyClassifierandCostRouting:FallbacktoOpus4.8inHigh-RiskDomains,CostsRoutedbyTaskComplexityAccordingtoAnthropicengineeringdocuments,Fable5hasabuilt-insafetyclassifierthatrefusestorespondandautomaticallyfallsbacktoOpus4.8inareassuchassecurityvulnerabilityresearch,biology,chemistry,andmodeldistillation;its319-pagesystemcarddocumentsthefullscopeoftheclassifier,withsomedowngradebehaviorsdiscovereddeepwithinthedocumentafteritsJune2026launch.ThecostroutingpatternactuallyusedbyAnthropicengineers:Fable5actsasorchestrator(multi-dayplanning,delegatingsub-agents,visualverification);Opus4.8handlesdifficultbutboundedsub-tasks(architecturedecisions,complexdebugging)andfallbackforclassifierblocks;Sonnet4.6handleshigh-volumeworkertasks(linting,simplerefactoring,documentationupdates);Haiku4.5servesasscorersub-agentandcheapclassifier.##FrequentlyAskedQuestions####**HowdoesFable5's"self-improvement"differfrom"self-learning"?**AccordingtoAnthropicengineeringdocuments,self-learningmeansthemodelupdatesitsownweightsbasedonwhatitlearns;Fable5doesnotdothis,andcurrentlynopubliclyavailablemodelhasachievedthiscapabilityinproduction.Self-improvementmeansthesystemaroundthemodelcompoundswitheachexecution:memoryaccumulatesverifiedfacts,Skillsbecomesharperbyincorporatingedgecases,evalloopsrefineprompts;themodelitselfremainsunchanged,buttheoperatingenvironmentbecomessharper.####**WhatareRoutines,andwhenwilltheylaunch?**AccordingtoAnthropicengineeringdocuments,RoutinesarestoredClaudeCodeconfigurations(prompts,repositories,connectors,permissions)thatexecuteonAnthropic-hostedcloudinfrastructureundertriggerconditions,evenwhenthelocalmachineisoff;RoutineslaunchedasaresearchpreviewonApril14,2026,supportingthreetriggertypes:scheduledtriggers,APItriggers,andGitHubeventtriggers.####**Whyisanindependentverifiersub-agentbetterthanself-critique?**AccordingtoAnthropicengineerPrithviRajasekaran'sengineeringblogpostandFable5launchdata,whenamodelevaluatesitsownoutput,itseesitsownreasoningtraceandtendstoagreewithconclusionsitpreviouslywrote;anotheragentonlyseestheoutputandscoringcriteria.Theverifierhasnostakeintheproducer'sgame,canexplorealargerhypothesisspace,andrecoverfromnegativeintermediateresults.

Disclaimer: The information on this page may come from third-party sources and is for reference only. It does not represent the views or opinions of Gate and does not constitute any financial, investment, or legal advice. Virtual asset trading involves high risk. Please do not rely solely on the information on this page when making decisions. For details, see the Disclaimer.