從工具到自主化，構(gòu)建更強(qiáng)大的Agent系統(tǒng)范文棟CAMEL-AIAgentfrom1986thereisnotrick.Thepowerofintelligencestemsfromourvastdiversity,notfromanysingle,perfectprinciple.AgentsTheThelearneranddecision-makeriscalledtheAgentAgentAgentinteractswithanenvironmentthroughactionsEnvironmentEnvironmentpresentsnewstatesandrewardsTheagent-environmentinteractioninRLAgentsAutomatedVehiclesGameAutomatedVehiclesGameAgentLanguageModelsasAgentsKeydifferences:●LanguageasInput●LanguageasOutput●StateandActionareexpressedasnaturallanguage●LilianWeng:https://lilianweng.github.io/posts/2023-06-23-agent/LanguageModelsasAgentsKeyFeatures:-Memory:Manageschathistoryandcontextwindow-Tools:Supportsbothinternalandexternalfunctioncalls-StepLoop:HandletaskrequiremultiplerequestwithonestepAgentsclassTaskPrioritiz)●TaskDecomposition●Self-ReflectionAgentsmeta_dict=dict(critic_criteria='Helpbetteraccomplisys_msg=sys_msg_gen().from_dictcritic_agent=CriticAgent(print(critic_agent.systask:{task}.YourjobistoselectanoptionfromtheirproposalsandYourselectioncriteriaareHelpbetteraccompl●TaskDecomposition●Self-ReflectionThenextscalinglaws?#Parameters->#Agents?●Data->Environments?●Training->Memory/Interaction?TheneweraofMulti-agentsystemwithLLMagents?Multi-AgentFrameworkTheFirstBabyAGIBabyAGIMetaGPT >ChatChatDevAutoGPTCrewAICAMEL●Idea●Roleassignment●Taskagents●ChatagentsRole-PlayingFrameworkCAMEL(NeurIPS2023):/abs/2303.17760CAMEL●Idea●Roleassignment●Taskagents●ChatagentsCAMEL(NeurIPS2023):/abs/2303.17760CAMELWorkforce●HierarchicalArchitecture●TaskPlanningandDecomposition●TaskChannelCommunication●RobustFailureHandlingCAMEL(NeurIPS2023):/abs/2303.17760MoreAgentsBetweenthanone?●CAMELagentsarebetteroneagent>70%on200tasksAgentEvaluationResults●GPT-4evaluationalignswithHumanevaluationAgentEvaluationResultsCAMEL(NeurIPS2023):/abs/2303.17760OWL:OptimizedWorkforceLearningOWL:OptimizedWorkforceLearning●58.18averagescoreonGAIAbenchmark●Ranking#1amongopen-sourceframeworks!CRAB:Cross-environmentAgentBenchmarkForMultimodalLanguageModelAgentsCRAB:Cross-environmentAgentBenchmarkforMultimodalLanguageModelAgents(NeurIPS2024OWAworkshop):/pdf/2407.01511CRABAgentsandEnvironmentsphoneCrab:Cross-environmentAgentBenchmarkforMultimodalLanguageModelAgentsOASIS:SimulateSocialMediawith1millionAgents●Upto1millionagents●Replicatesocialscienceexperiments●ExploredynamicofagentsocietyOASIS:OpenAgentSocialInteractionSimulationswithOneMillionAgents(NeurIPS2024OWAworkshop):/abs/2411.11581AgentswithGraphRAGAgentswithGraphRAGRole-PlayingScraperforReport&KnowledgeGraphGeneration●AutonomousTaskSolving●KnowledgeGeneration●ScalableSolutionAgentswithGraphRAGRole-PlayingScraperforReport&KnowledgeGraphGeneration●AutonomousTaskSolving●KnowledgeGeneration●ScalableSolutionAgentswithGraphRAGRole-PlayingScraperforReport&KnowledgeGraphGeneration●AutonomousTaskSolving●KnowledgeGeneration●ScalableSolutionAgenticDataGeneration/cookbooks/sft_data_generation_and_unsloth_finetuning_tinyllama.htmlAgenticDataGeneration●AISociety:50assistantroles,50userrolesand10tasks●Code:20programminglanguages,50domains,and50tasksAgenticDataGeneration●GeneratedatafromGPT3.5/4●FinetuneLlamamodelsAgenticDataGeneration●180+modelsonHuggingface●MosaicMLMPT-30B-Chat●HuggingFaceM4/idefics2-8bTheNextScalingLaws?F