REEFContinual learning infrastructure for self-improving agentsServes agent requests, records feedback, trains updates, publishes versionsgithub.com/Human-Agent-Society/reef1 / 3Self-evolving agents must learn from the experience they generate at inference time.Both the model and the harness need updates.REAL USEusers and appsrequestsresponsesAGENTv1modelweightsharnessprompts · skills · toolsmemory · orchestration1trajectoriesexecution resultsuser feedbackexperience2UPDATEgrow3model trainerRL · distillationharness trainerproposed editsevaluate, then publishAGENTv2modelweightsharnessprompts · skills · toolsmemory · orchestration41ServePOST /v1/chat/completions2ObservePOST /reef/report3Growtrain weights or update harness4Commitevaluate, then publishaccepted updates are served to the next request
REEFSystem designArchitecture and supported recipesgithub.com/Human-Agent-Society/reef2 / 3ArchitectureENTRYPOINTSmethod plug-insrecipes/*HTTP · CLI · reef-clientreef/serviceKERNELapplication kernelreef/dispatcher · reef/scenario: coordination, commit ordering, recoveryDOMAINSservingruntime · surfaceSGLang · vLLMevolutiontrain · harnessSlime · harness adaptersstateartifact · recordsGit LFS · SQLiteADAPTERSModules for each step of the loop1Serveservice · runtime2Observerecords · train/processors3Growrecipe · train4Committrain/evaluation · artifact · surfaceSupported recipesmodel weightstrains the model with Slime and SGLangonline RL from live trafficSAOmath reasoning · IMO-BenchOpenClaw-RLpersonalized chattest-time trainingTTT-Discoveropen-ended discoveryGuidance-TTTprogram searchagent harnessworks with any model endpoint, no training GPUprompt evolutionGEPAmath reasoning · AIME 2025skill evolutionSkillClawagent tasks · WildClawBenchcomposition searchMeta-Harnessterminal tasks · Terminal-Bench
REEFWhat Reef can do nextFour research directionsgithub.com/Human-Agent-Society/reef3 / 3DAY · SERVE, RECORDNIGHT · REPLAY, GENERATE, TRAINAGENTv1modelharnessexperiencegenerated tasksreplays and variationsAGENTv2modelharnesstrain while idle1Evolve during sleepGenerate training tasks from the day's trafficand train while the agent is idle.ORGANIZATIONshared experience · shared versionsPLANNERmodelharnessCODERmodelharnessREVIEWERmodelharnessOPERATORmodelharnesshand-off2Multi-agentMultiple agents share one learning loop, sotheir roles and hand-offs are updated together.EXPERIENCEGATElearnrefuseattributed · verifiedstale · unverifiableweightsharnessv1v2v1v2v3v4co-evolve3Control over evolutionDecide which experience to learn from, andwhen to update the weights or the harness.AGENTmodelharnessbest known?personalizationone user's preferencesscientific discoverya better boundsocial forecastinga future eventFIXED BENCHMARKREAL OUTCOMES4Open-ended tasksPersonalization, scientific discovery andsocial forecasting, evaluated on real outcomes.